TECHNOLOGY

RADAR
An opinionated guide
to technology frontiers

Vol.22
thoughtworks.com/radar
#TWTechRadar
Contributors
The Technology Advisory Board (TAB) is a group of 20 or so senior technologists at
ThoughtWorks. The TAB meets twice a year face-to-face and bi-weekly by phone. Its
primary role is to be an advisory group for for ThoughtWorks CTO, Rebecca Parsons.

The TAB acts as a broad body that can look at topics that affect technology and
The Technology Radar is prepared by the technologists at ThoughtWorks. We usually create the Radar at face-to-face meetings, but
given the global pandemic we’ve been living through, this is the first Technology Radar to
ThoughtWorks Technology Advisory Board be created via a virtual event.

Rebecca Martin Fowler Bharani Birgitta Camilla Erik Evan Fausto Hao Ian
Parsons (CTO) (Chief Scientist) Subramaniam Böckeler Crispim Dörnenburg Bottcher de la Torre Xu Cartwright

James Jonny Lakshminarasimhan Mike Neal Ni Rachel Scott Shangqi Zhamak

Lewis LeRoy Sudarshan Mason Ford Wang Laycock Shaw Liu Dehghani

TECHNOLOGY RADAR | 2
© ThoughtWorks, Inc. All Rights Reserved.
About
the Radar
ThoughtWorkers are passionate about
technology. We build it, research it, test it, open
source it, write about it, and constantly aim to
improve it — for everyone. Our mission is to
champion software excellence and revolutionize
IT. We create and share the ThoughtWorks
Technology Radar in support of that mission.
The ThoughtWorks Technology Advisory
Board, a group of senior technology leaders at
ThoughtWorks, creates the Radar. They meet
regularly to discuss the global technology
strategy for ThoughtWorks and the technology
trends that significantly impact our industry.

The Radar captures the output of the

Technology Advisory Board’s discussions in a
format that provides value to a wide range of
stakeholders, from developers to CTOs. The
content is intended as a concise summary.

We encourage you to explore these technologies.

The Radar is graphical in nature, grouping items
into techniques, tools, platforms and languages
& frameworks. When Radar items could appear
in multiple quadrants, we chose the one that
seemed most appropriate. We further group
these items in four rings to reflect our current
position on them.

For more background on the Radar, see

thoughtworks.com/radar/faq.

TECHNOLOGY RADAR | 3
© ThoughtWorks, Inc. All Rights Reserved.
 New

Radar at Moved in/out

a glance
No change

Our Radar is forward looking. To make

room for new items, we fade items that
haven’t moved recently, which isn’t a
reflection on their value but rather on
The Radar is all about tracking interesting our limited Radar real estate.
things, which we refer to as blips. We organize
the blips in the Radar using two categorizing
elements: quadrants and rings. The quadrants
represent different kinds of blips. The rings
indicate what stage in an adoption lifecycle we
think they should be in.
Hold Assess Trial Adopt Adopt Trial Assess Hold
A blip is a technology or technique that
plays a role in software development. Blips Adopt
are things that are ‘in motion’ — that is we
find their position in the Radar is changing We feel strongly that the industry should
— usually indicating that we’re finding be adopting these items. We use them
increasing confidence in them as they move when appropriate on our projects.
through the rings.
Trial
Worth pursuing. It’s important to understand how
to build up this capability. Enterprises can try this
technology on a project that can handle the risk.

Assess
Worth exploring with the goal of
understanding how it will affect your enterprise.

Hold
Proceed with caution.

TECHNOLOGY RADAR | 4
© ThoughtWorks, Inc. All Rights Reserved.
Themes for this edition
The Elephant in the Zoom X is Software Too Data Perspectives Maturing Kubernetes & Co. Cambrian
and Expanding Explosion
“Necessity is the mother of invention” We often encourage other parts of the
— Proverb software delivery ecosystem to adopt A theme that spanned many blips and As Kubernetes continues to consolidate its
beneficial engineering practices pioneered quadrants in this edition concerned market dominance, the inevitable supporting
Many companies have experimented by agile software development teams; we maturity in data, particularly techniques ecosystem thrives. We discussed a number
with the idea of remote working as the return to this topic so often because we and tools surrounding analytical data of blips surrounding Kubernetes in the
technology to enable it has slowly matured. keep finding niches where we see slow and machine learning. We note many tools, platforms and techniques quadrants,
But suddenly, a global pandemic has forced progress on this advice. For this Radar, we continuing innovations in the natural showing just how pervasive this subject
companies all over the world to rapidly decided to call out again infrastructure as language processing (NLP) space. We also has become. For example, Lens and k9s
and fundamentally change their way of code as well as pipelines as code, and we welcome both the emergence and continuing simplify cluster management, kind helps with
working to preserve some productivity. As also had a number of conversations about maturity of full-lifecycle machine learning local testing and Gloo offers an alternative
many have observed, “working from home” infrastructure configurations, ML pipelines tool suites, combining enduring engineering API Gateway. Hydra is an OAuth server
is starkly different from “being forced to and other related areas. We find that the practices with combinations of tools that optimized to run on Kubernetes, and Argo
work from home during a pandemic,” and teams who commonly own these areas do work well in an iterative manner, showing CD uses Kuberenetes’ native desired-state
we think there will be a journey ahead to not embrace enduring engineering practices that “machine learning is software too.” management to implement a CD server.
become fully productive in this new context. such as applying software design principles, Finally, for distributed architectures such as These developments indicate Kubernetes
automation, continuous integration, testing, microservices, we see great interest in data is perfectly poised to create a supporting
We’ve never believed that creating a Radar and so on. We understand that many mesh as a way to effectively serve and use ecosystem; it offers critical capabilities but
remotely was possible, and yet here we are factors hamper fast movement for some analytical data at scale in distributed systems. with abstractions that are often too low
— this is the first Radar we’ve ever produced engineering practices: complexity (both As the industry thinks more diligently about level or advanced for most users. Thus,
without meeting in person. Many of the essential and accidental), lack of knowledge, how data should work in modern systems, the complexity void fills with tooling to
proposed blips spoke to the pressing need political impediments, lack of suitable tooling we’re encouraged by the general direction either ease the configuration and the use
to enable first-class remote collaboration. and many others. However, the benefits to and opening perspectives in this arena and of Kubernetes or supply something missing
We didn’t want to ignore the elephant in the organizations that embrace agile software expect to see exciting innovations in the near from the core functionality. As Kubernetes
room and not comment on the crisis, but delivery practices are clear and worth some future. continues to dominate, we see a rich
doing a good job of remote-first collaboration effort to achieve. ecosystem growing and expanding to take
is a deep and nuanced subject and certainly advantage of its strengths and address its
not all of our advice would fit in the Radar weaknesses. As this ecosystem matures,
format. So alongside this edition you’ll find a we expect it to evolve toward a new set of
podcast where we discuss our experiences higher-level abstractions offering the benefits
in creating the Radar remote-first, a written of Kubernetes without the bewildering range
experience report including advice on of options.
remote-first productivity, a webinar covering
tech strategies in a crisis and links to other
ThoughtWorks materials, including our
remote working playbook. We hope that
these materials, together with other internet
resources, will help organizations that attempt
to navigate these unknown waters.

TECHNOLOGY RADAR | 5
© ThoughtWorks, Inc. All Rights Reserved.
 Adopt

The Radar Adopt

Tools
Techniques

1. Applying product management to internal 50. Cypress

platforms 51. Figma
2. Infrastructure as code
3. Micro frontends Trial
4. Pipelines as code 52. Dojo
5. Pragmatic remote pairing 53. DVC
6. Simplest possible feature toggle 54. Experiment tracking tools
for machine learning
Trial 55. Goss
7. Continuous delivery for machine learning (CD4ML) 56. Jaeger
8. Ethical bias testing 27 57. k9s
9. GraphQL for server-side resource aggregation 58. kind
10. Micro frontends for mobile 26 59. mkcert
11. Platform engineering product teams 60. MURAL
12. Security policy as code 22 68 61. Open Policy Agent (OPA)
67
13. Semi-supervised learning loops 62. Optimal Workshop
14. Transfer learning for NLP 21 69 63. Phrase
15. Use “remote native” processes and approaches 70 64. ScoutSuite
25
16. Zero trust architecture (ZTA) 65. Visual regression testing tools
20 66. Visual Studio Live Share
Assess 52 71
54
15
17. Data mesh Assess
16 55 72
18. Decentralized identity 13 53 67. Apache Superset
24 19 14
19. Declarative data pipeline definition 58 68. AsyncAPI
20. DeepWalk 12 57
73
69. ConfigCat
21. Managing stateful systems via container 56 59 70. Gitpod
orchestration 18 71. Gloo
22. Preflight builds 11 74
10 72. Lens
6 60 62
61 73. Manifold
Hold 23 5 74. Sizzy
9 63 75
23. Cloud lift and shift 75. Snowpack
24. Legacy migration feature parity 17
8
50 76. tfsec
25. Log aggregation for business analytics 3 4 51 64 76
65
26. Long-lived branches with Gitflow 7
1
2
Hold
66
27. Snapshot testing only
Hold Assess Trial Adopt Adopt Trial Assess Hold

28
Adopt 89 Adopt

Languages & Frameworks

Platforms

39 30 79
28. .NET Core 97
77. React Hooks
29. Istio 87 88 78. React Testing Library
31 78
79. Vue.js
Trial 40 77
32 29 96
30. Anka Trial
31. Argo CD 41 86 80. CSS-in-JS
33
32. Crowdin 34 85 81. Exposed
33. eBPF 83 82. GraphQL Inspector
84 95
34. Firebase 42 83. Karate
35. Hot Chocolate 35 81 84. Koin
36. Hydra 43 37 94 85. NestJS
37. OpenTelemetry 36
38
80 82 86. PyTorch
38. Snowflake 87. Rust
44 88. Sarama
93
Assess 92
89. SwiftUI
39. Anthos 49 45 98
40. Apache Pulsar 46 47 48 91 Assess
41. Cosmos 90. Clinic.js Bubbleprof
42. Google BigQuery ML 90 91. Deequ
43. JupyterLab 92. ERNIE
44. Marquez 93. MediaPipe
45. Matomo 94. Tailwind CSS
46. MeiliSearch 95. Tamer
47. Stratos 96. Wire
48. Trillian 97. XState

Hold Hold
49. Node overload 98. Enzyme

New Moved in/out No change

TECHNOLOGY RADAR Vol. 22

Techniques
Techniques 26
27
Adopt
1. Applying product management
to internal platforms
2. Infrastructure as code
3. Micro frontends
Applying product management 4. Pipelines as code
22 68
to internal platforms 67 5. Pragmatic remote pairing
Adopt 21 6. Simplest possible feature
69
70 toggle
25
More and more companies are building
internal platforms to roll out new digital Trial
20 52
solutions quickly and efficiently. Companies 54 71
7. Continuous delivery for
that succeed with this strategy are applying 15
16
machine 72 learning (CD4ML)
product management to internal platforms. 13 53 55
24 19 14
8. Ethical bias testing
This means establishing empathy with 58
12 57
9. GraphQL for server-side
internal consumers (the development teams) resource aggregation
73
56
and collaborating with them on the design. 59 10. Micro frontends for mobile
Platform product managers create roadmaps 18
11 11. Platform engineering
74 product
and ensure the platform delivers value to 10
6 60 teams
62
the business and enhances the developer 61
23 5 12. Security policy as code
experience. Unfortunately, we’re also 9 63
13. Semi-supervised 75learning loops
seeing less successful approaches, where 50 14. Transfer learning for NLP
17
teams create a platform in the void, based 8 3 4 51 15. Use 65
64 “remote native” 76 processes
on unverified assumptions and without 1 and approaches
internal customers. These platforms, often 7 2 66
16. Zero trust architecture (ZTA)
despite aggressive internal tactics, end
up being underutilized and a drain on the Hold Assess Trial Adopt Adopt Trial Assess Hold
Assess
organization’s delivery capability. As usual, 17. Data mesh
28 89
good product management is all about 39 30 79 18. Decentralized identity
building products that consumers love. 19. Declarative data pipeline
97
maintainability, and using automated allowed teams to scale the delivery of 87 88
31 78 definition
testing and deployment are all critical independently deployed and maintained
40 20. DeepWalk
Infrastructure as code practices. Those of us with a deep software services. Unfortunately, we’ve also seen 77
32 29 21. Managing stateful96 systems via
Adopt and infrastructure background need to many teams create a front-end monolith —
41 86 container orchestration
empathize with and support colleagues a large, entangled browser application that
33
22. Preflight builds
Although infrastructure as code is a who do not. Saying “treat infrastructure like sits on 34
top of the back-end services — largely 85
code” isn’t enough; we need to ensure the neutralizing the benefits of microservices. 83
relatively old technique (we’ve featured 84 Hold 95
42 world
hard-won learnings from the software Micro frontends have continued to gain in
it in the Radar in 2011), it has become 35 23. Cloud lift and shift
vitally important in the modern cloud era are also applied consistently throughout the43 popularity since they were first introduced. 81
37 24. Legacy 94
migration feature parity
where the act of setting up infrastructure infrastructure realm. We’ve seen many teams adopt some form
36 80 82 25. Log aggregation for business
has become the passing of configuration of this architecture as a38
way to manage
analytics
instructions to a cloud platform. When we the complexity of multiple developers
44 26. Long-lived branches with
say “as code” we mean that all the good Micro frontends and teams contributing to the same user 93
Gitflow
92
practices we’ve learned in the software Adopt 49 experience. In June of last year, one of the
45 98
27. Snapshot testing only
world should be applied to infrastructure. originators of this
47technique published an
46 48 91
Using source control, adhering to We’ve seen significant benefits from introductory article that serves as a reference
the DRY principle, modularization, introducing microservices, which have for micro frontends. It shows how this style90

8 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Techniques
can be implemented using various web in this situation we recommend pragmatic is more than just training models and
programming mechanisms and builds out remote pairing: adjusting pairing practices serving them. It requires implementing end-
an example application using React.js. We’re to what’s possible given the tools at hand. to-end and continuously repeatable cycles
confident this style will grow in popularity Consider tools such as Visual Studio Live of training, testing, deploying, monitoring
as larger organizations try to decompose UI Share for efficient, low-latency collaboration. and operating the models. Continuous
We firmly believe in pair development across multiple teams. Only resort to pixel-sharing if both participants delivery for machine learning (CD4ML) is
programming. But in a reside in relative geographic proximity and a technique that enables reliable end-to-
post COVID-19, remote- have high-bandwidth internet connections. end cycles of development, deploying and
working world, successful Pipelines as code Pair developers who are in similar time zones monitoring machine learning models. The
pairing requires a healthy Adopt rather than expecting pairing to work between underpinning technology stack to enable
participants regardless of their location. If CD4ML includes tooling for accessing
dose of pragmatism to be The pipelines as code technique emphasizes pairing isn’t working for logistical reasons, and discovering data, version control of
effective. that the configuration of delivery pipelines fall back to practices such as individual artefacts (such as data, model and code),
that build, test and deploy our applications programming augmented via code reviews, continuous delivery pipelines, automated
(Pragmatic remote pairing) or infrastructure should be treated as pull-request collaboration (but beware long- environment provisioning for various
code; they should be placed under source lived branches with Gitflow) or shorter pairing deployments and experiments, model
control and modularized in reusable sessions for critical parts of the code. We’ve performance assessment and tracking, and
We encourage you components with automated testing and engaged in remote pairing for years, and model operational observability. Companies
to use the simplest deployment. As organizations move to we’ve found it to be effective if done with a can choose their own tool set depending on
possible feature toggle decentralized autonomous teams building dose of pragmatism. their existing tech stack. CD4ML emphasizes
microservices or micro frontends, the need automation and removing manual handoffs.
instead of unnecessarily
for engineering practices in managing CD4ML is our de facto approach for
complex feature toggle pipelines as code increases to keep building
frameworks. Simplest possible feature toggle developing ML models.
and deploying software consistent within Adopt
the organization. This need has given rise
(Simplest possible feature to delivery pipeline templates and tooling
toggle)
Unfortunately, feature toggles are less Ethical bias testing
that enable a standardized way to build and common than we’d like, and quite often Trial
deploy services and applications. Such tools we see people mixing up its types and use
use the declarative delivery pipelines of cases. It’s quite common to come across Over the past year, we’ve seen a shift in
applications, adopting a pipeline blueprint teams that use heavyweight platforms interest around machine learning and deep
to execute the underlying tasks for various such as LaunchDarkly to implement neural networks in particular. Until now,
stages of a delivery lifecycle such as build, feature toggles, including release toggles, tool and technique development has been
test and deployment; and they abstract away to benefit from Continuous Integration, driven by excitement over the remarkable
implementation details. The ability to build, when all you need are if/else conditionals. capabilities of these models. Currently,
test and deploy pipelines as code should be Therefore, unless you need A/B testing though, there is rising concern that these
one of the evaluation criteria for choosing a or canary release or hand over feature models could cause unintentional harm.
CI/CD tool. release responsibility to business folks, we For example, a model could be trained
encourage you to use the simplest possible inadvertently to make profitable credit
feature toggle instead of unnecessarily decisions by simply excluding disadvantaged
Pragmatic remote pairing complex feature toggle frameworks. applicants. Fortunately, we’re seeing a
Adopt growing interest in ethical bias testing that
will help to uncover potentially harmful
We firmly believe that pair programming Continuous delivery for machine decisions. Tools such as lime, AI Fairness 360
improves the quality of code, spreads or What-If Tool can help uncover inaccuracies
learning (CD4ML)
knowledge throughout a team and allows that result from underrepresented groups
Trial
overall faster delivery of software. In a post in training data and visualization tools such
COVID-19 world, however, many software Applying machine learning to make the as Google Facets or Facets Dive can be
teams will be distributed or fully remote, and business applications and services intelligent used to discover subgroups within a corpus

TECHNOLOGY RADAR | 9
© ThoughtWorks, Inc. All Rights Reserved.
 Techniques
of training data. We’ve used lime (local presents the challenge of maintaining team disruption. For example, access control
interpretable model-agnostic explanations) autonomy while integrating their work into policies define and enforce who can access
in addition to this technique in order to a single app. Although we’ve seen teams which services and resources under what
understand the predictions of any machine- writing their own frameworks to enable this circumstances; or network security policies
learning classifier and what classifiers (or development style, existing modularization can dynamically limit the traffic rate to a
models) are doing. frameworks such as Atlas and Beehive can particular service. The complexity of the There is rising concern
also simplify the problem of integrating technology landscape today demands that some machine-
multiteam app development. treating security policy as code: define learning models could
GraphQL for server-side and keep policies under version control, cause unintentional harm.
resource aggregation automatically validate them, automatically Fortunately, we’re seeing a
Trial Platform engineering product deploy them and monitor their
growing interest in ethical
teams performance. Tools such as Open Policy
We see more and more tools such as Apollo Agent or platforms such as Istio provide bias testing that will help
Trial
Federation that can aggregate multiple flexible policy definition and enforcement to uncover potentially
GraphQL endpoints into a single graph. The adoption of cloud and DevOps — while mechanisms that support the practice of harmful decisions.
However, we caution against misusing increasing the productivity of teams who security policy as code.
GraphQL, especially when turning it into a can now move more quickly with reduced (Ethical bias testing)
server-to-server protocol. Our practice is dependency on centralized operations teams
to use GraphQL for server-side resource and infrastructure — also has constrained Semi-supervised learning loops
aggregation only. When using this pattern, teams that lack the skills to self-manage a Trial
the microservices continue to expose well- full application and operations stack. Some Micro frontends have
defined RESTful APIs, while under-the-hood organizations have tackled this challenge by Semi-supervised learning loops are a class been widely adopted for
aggregate services or BFF (Backend for creating platform engineering product teams. of iterative machine-learning workflows that Web UIs. Now we’re seeing
Frontends) patterns use GraphQL resolvers These teams maintain an internal platform take advantage of the relationships to be
this architectural style
as the implementation for stitching that enables delivery teams to deploy and found in unlabeled data. These techniques
may improve models by combining labeled
come to mobile too.
resources from other services. The shape operate systems with reduced lead time and
of the graph is driven by domain-modeling stack complexity. The emphasis here is on API- and unlabeled data sets in various ways. In
other cases they compare models trained (Micro frontends for mobile)
exercises to ensure ubiquitous language driven self-service and supporting tools, with
is limited to subgraphs where needed (in delivery teams still responsible for supporting on different subsets of the data. Unlike
the case of one-microservice-per-bounded- what they deploy onto the platform. either unsupervised learning where a
context). This technique simplifies the Organizations that consider establishing such machine infers classes in unlabeled data or
internal implementation of aggregate a platform team should be very cautious not supervised techniques where the training
services or BFFs, while encouraging good to accidentally create a separate DevOps set is entirely labeled, semi-supervised
modeling of services to avoid anemic REST. team, nor should they simply relabel their techniques take advantage of a small set
existing hosting and operations structure of labeled data and a much larger set of
as a platform. If you’re wondering how to unlabeled data. Semi-supervised learning
Micro frontends for mobile best set up platform teams, we’ve been is also closely related to active learning
Trial using the concepts from Team Topologies techniques where a human is directed to
to split platform teams in our projects into selectively label ambiguous data points.
Since introducing it in the Radar in 2016, enablement teams, core “platform within a Since expert humans that can accurately
we’ve seen widespread adoption of micro platform” teams and stream-focused teams. label data are a scarce resource and
frontends for web UIs. Recently, however, labeling is often the most time-consuming
we’ve seen projects extend this architectural activity in the machine-learning workflow,
style to include micro frontends for mobile Security policy as code semi-supervised techniques lower the cost
applications as well. When the application Trial of training and make machine learning
becomes sufficiently large and complex, feasible for a new class of users. We’re also
it becomes necessary to distribute the Security policies are rules and procedures seeing the application of weakly supervised
development over multiple teams. This that protect our systems from threats and techniques where machine-labeled data

10 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Techniques
is used but is trusted less than the data and “everybody always remote” meetings age-old assumption that we must centralize
labeled by humans. are other approaches our teams practice by big analytical data to use it, have data all in
default to optimize for location fluidity. one place or be managed by a centralized
data team to deliver value. Data mesh
Transfer learning for NLP claims that for big data to fuel innovation,
Zero trust architectures Trial Zero trust architecture (ZTA) its ownership must be federated among
stress the importance of Trial domain data owners who are accountable
securing all access and We had this technique in Assess previously. for providing their data as products (with
communications and The innovations in the NLP landscape The technology landscape of organizations the support of a self-serve data platform to
today is increasingly more complex with abstract the technical complexity involved in
enforcing policies as continue at a great pace, and we’re able to
leverage these innovations in our projects assets — data, functions, infrastructure and serving data products); it must also adopt a
code based on the least users — spread across security boundaries,
thanks to the ubiquitous transfer learning new form of federated governance through
privilege. for NLP. The GLUE benchmark (a suite such as local hosts, multiple cloud providers automation to enable interoperability
of language understanding tasks) scores and a variety of SaaS vendors. This demands of domain-oriented data products.
(Zero trust architecture (ZTA)) have seen dramatic progress over the past a paradigm shift in enterprise security Decentralization, along with interoperability
couple of years with average scores moving planning and systems architecture, moving and focus on the experience of data
from 70.0 at launch to some of the leaders from static and slow-changing security policy consumers, are key to the democratization of
crossing 90.0 as of April 2020. A lot of our management, based on trust zones and innovation using data.
projects in the NLP domain are able to network configurations, to dynamic, fine-
make significant progress by starting from grained security policy enforcement based on If your organization has a large number of
pretrained models from ELMo, BERT, and temporal access privileges. domains with numerous systems and teams
ERNIE, among others, and then fine-tuning generating data or a diverse set of data-driven
them based on the project needs. Zero trust architecture (ZTA) is an use cases and access patterns, we suggest
organization’s strategy and journey to you assess data mesh. Implementation of
implement zero-trust security principles data mesh requires investment in building a
Use “remote native” processes for all of their assets — such as devices, self-serve data platform and embracing an
infrastructure, services, data and users — organizational change for domains to take
and approaches
and includes implementing practices such on the long-term ownership of their data
Trial
as securing all access and communications products, as well as an incentive structure that
Distributed teams come in many shapes and regardless of the network location, enforcing rewards domains serving and utilizing data as
setups; delivery teams in a 100% single-site policies as code based on the least privilege a product.
co-located setup, however, have become the and as granular as possible, and continuous
exception for us. Most of our teams are either monitoring and automated mitigation of
multisite teams or have at least some team threats. Our Radar reflects many of the Decentralized identity
members working off-site. Therefore, using enabling techniques such as security policy Assess
“remote native” processes and approaches by as code, sidecars for endpoint security and
default can help significantly with the overall BeyondCorp. If you’re on your journey toward Since the birth of the internet, the technology
team flow and effectiveness. This starts with ZTA, refer to the NIST ZTA publication to learn landscape has experienced an accelerated
making sure that everybody has access to more about principles, enabling technology evolution toward decentralization. While
the necessary remote systems. Moreover, components and migration patterns as well as protocols such as HTTP and architectural
using tools such as Visual Studio Live Share, Google’s publication on BeyondProd. patterns such as microservices or data mesh
MURAL or Jamboard turn online workshops enable decentralized implementations,
and remote pairing into routines instead of identity management remains centralized. The
ineffective exceptions. But “remote native” Data mesh emergence of distributed ledger technology
goes beyond a lift-and-shift of co-location Assess (DLT), however, provides the opportunity to
practices to the digital world: Embracing more enable the concept of decentralized identity. In
asynchronous communication, even more Data mesh is an architectural and a decentralized identity system, entities — that
discipline around decision documentation, organizational paradigm that challenges the is, discrete identifiable units such as people,

TECHNOLOGY RADAR | 11
© ThoughtWorks, Inc. All Rights Reserved.
 DeepWalk
Techniques
organizations and things — are free to use any the CI on a master branch can be ineffective
shared root of trust. In contrast, conventional Assess if the team is too big, the builds are slow or
identity management systems are based on flaky, or the team lacks the discipline to run
centralized authorities and registries such DeepWalk is an algorithm that helps the full test suite locally. In this situation a red
as corporate directory services, certificate apply machine learning on graphs. When build can block multiple devs or pairs of devs.
authorities or domain name registries. working on data sets that are represented Instead of fixing the underlying root cause — The foundation of BitCoin
as graphs, one of the key problems is slow builds, the inability to run tests locally — distributed ledger
The development of decentralized identifiers to extract features from the graph. This or monolithic architectures that necessitate technology (DLT) — is
— globally unique, persistent and self- is where DeepWalk can help. It uses many people working in the same area enabling the emergence of
sovereign identifiers that are cryptographically SkipGram to construct node embeddings — teams usually rely on feature branches decentralized identities.
verifiable — is a major enabling standard. by viewing the graph as a language to bypass these issues. We discourage
Although scaled implementations of where each node is a unique word in feature branches, given they may require
(Decentralized identity)
decentralized identifiers in the wild are the language and random walks of finite significant effort to resolve merge conflicts,
still rare, we’re excited by the premise of length on the graph constitutes a sentence. and they introduce longer feedback loops
this movement and have started using the These embeddings can then be used by and potential bugs during conflict resolution.
concept in our architecture. For the latest various ML models. DeepWalk is one of Instead, we propose using preflight builds as When working on data
experiments and industry collaborations, the techniques we’re trialling on some of an alternative: these are pull request–based sets that are represented
check out Decentralized Identity Foundation. our projects where we’ve needed to apply builds for “micro branches” that live only for as graphs, one of the key
machine learning on graphs. the duration of the pipeline run, with the problems is to extract
branch opened for every commit. To help features from the graph.
Declarative data pipeline automate this workflow, we’ve come across This is where DeepWalk
definition Managing stateful systems via bots such as Bors, which automates merging
can help.
Assess container orchestration to master and branch deletion in case the
Assess mini branch build succeeds. We’re assessing
(DeepWalk)
Many data pipelines are defined in a large, this flow, and you should too; but don’t use
more or less imperative script written in We recommend caution in managing this to solve the wrong problem, as it can
Python or Scala. The script contains the stateful systems via container orchestration lead to misuse of branches and may cause
logic of the individual steps as well as the platforms such as Kubernetes. Some more harm than benefit.
code chaining the steps together. When databases are not built with native support
faced with a similar situation in Selenium for orchestration — they don’t expect a
tests, developers discovered the Page scheduler to kill and relocate them to a Cloud lift and shift
Object pattern, and later many behavior- different host. Building a highly available Hold
driven development (BDD) frameworks service on top of such databases is not
implemented a split between step trivial, and we still recommend running It is rather curious, that after over a decade
definitions and their composition. Some them on bare metal hosts or a virtual of industry experience with cloud migration,
teams are now experimenting with bringing machine (VM) rather than to force-fit them we still feel it’s necessary to call out cloud
the same thinking to data engineering. into a container orchestration platform. lift and shift, a practice that views the cloud
A separate declarative data pipeline simply as a hosting solution, resulting in
definition, maybe written in YAML, contains the replication of an existing architecture,
only the declaration and sequence of steps. Preflight builds security practices and IT operational models
It states input and output data sets but Assess in the cloud. This fails to realize the cloud’s
refers to scripts if and when more complex promises of agility and digital innovation. A
logic is needed. With A La Mode, we’re Even though we strongly advocate in favor cloud migration requires intentional change
seeing the first open source tool appear in of CI rather than Gitflow, we know that across multiple axes toward a cloud-native
this space. committing straight to the trunk and running state, and depending on the unique migration

12 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Techniques Long-lived branches with Gitflow
circumstances, each organization might Standish Group report) and business
end up somewhere on the spectrum from processes that have evolved over time. Hold
cloud lift and shift to cloud native. Systems Replacing these features is a waste. Our
architecture, for example, is one of the advice: Convince your customers to take a Five years ago we highlighted the problems
pillars of delivery agility and often requires step back and understand what their users with long-lived branches with Gitflow.
Logs intended for change. The temptation to simply lift and shift currently need and prioritize these needs Essentially, long-lived branches are the
technical observability existing systems as containers to the cloud against business outcomes and metrics opposite of continuously integrating all
are often inadequate can be strong. While this tactic can speed up — which often is easier said than done. changes to the source code, and in our
to infer deep customer cloud migration, it falls short when it comes This means conducting user research and experience continuous integration is the
understanding. to creating agility and delivering features applying modern product development better approach for most kinds of software
and value. Enterprise security in the cloud practices rather than simply replacing the development. Later we extended our caution
is fundamentally different from traditional existing ones. to Gitflow itself, because we saw teams using
(Log aggregation for business
perimeter-based security through firewalls it almost exclusively with long-lived branches.
analytics)
and zoning, and it demands a journey toward Today, we still see teams in settings where
zero trust architecture. The IT operating model Log aggregation for business continuous delivery of web-based systems
too has to be reformed to safely provide analytics is the stated goal being drawn to long-lived
Snapshot testing is cloud services through self-serve automated Hold branches. So we were delighted that the
undeniably useful when platforms and empower teams to take more author of Gitflow has now added a note to his
working with legacy of the operational responsibility and gain Several years ago, a new generation of log original article, explaining that Gitflow was not
systems. But it shouldn’t autonomy. Last but not least, organizations aggregation platforms emerged that were intended for such use cases.
be the primary test must build a foundation to enable continuous capable of storing and searching over vast
mechanism for such change, such as creating pipelines with amounts of log data to uncover trends and
continuous testing of applications and insights in operational data. Splunk was Snapshot testing only
systems.
infrastructure as a part of the migration. the most prominent but by no means the Hold
These will help the migration process, result in only example of these tools. Because these
(Snapshot testing only)
a more robust and well-factored system and platforms provide broad operational and The value of snapshot testing is undeniable
give organizations a way to continue to evolve security visibility across the entire estate of when working with legacy systems by
and improve their systems. applications, administrators and developers ensuring that the system continues to work
have grown increasingly dependent on them. and the legacy code doesn’t break. However,
This enthusiasm spread as stakeholders we’re seeing the common, rather harmful
Legacy migration feature parity discovered that they could use log aggregation practice of using snapshot testing only as
Hold for business analytics. However, business the primary test mechanism. Snapshot tests
needs can quickly outstrip the flexibility and validate the exact result generated in the
We find that more and more organizations usability of these tools. Logs intended for DOM by a component, not the component’s
need to replace aging legacy systems to keep technical observability are often inadequate behavior; therefore, it can be weak and
up with the demands of their customers to infer deep customer understanding. unreliable, fostering the “only delete the
(both internal and external). One antipattern We prefer either to use tools and metrics snapshot and regenerate it” bad practice.
we keep seeing is legacy migration feature designed for customer analytics or to take a Instead, you should test the logic and
parity, the desire to retain feature parity more event-driven approach to observability behavior of the components emulating what
with the old. We see this as a huge missed where both business and operational events users would do. This mindset is encouraged
opportunity. Often the old systems have are collected and stored in a way they can be by tools in the Testing Library family.
bloated over time, with many features replayed and processed by more purpose-
unused by users (50% according to a 2014 built tools.

TECHNOLOGY RADAR | 13
© ThoughtWorks, Inc. All Rights Reserved.
TECHNOLOGY RADAR Vol. 22

Platforms
 23 61
5
9 63 75

17 50
8 3 64
4 51 65 76
1
7 2 66

Platforms Hold Assess

39
Trial

30
Adopt

28
79
Adopt Trial
Adopt
28. .NET Core
29. Istio
89
Assess

97
Hold

87 88
Trial
78
.NET Core 31
30. Anka
Adopt 40
29
77 31. Argo CD
32 96
41 86 32. Crowdin
We previously had .NET Core in Adopt, 33 33. eBPF
indicating that it had become our default 34 85 34. Firebase
for .NET projects. But we felt it’s worth again 83
84 35. Hot Chocolate
95
calling attention to .NET Core. With the 42 36. Hydra
35 81
release of .NET Core 3.x last year, the bulk 43 37. OpenTelemetry
94
37
of the features from .NET Framework have 36 80 38. Snowflake
82
now been ported into .NET Core. With the 38
announcement that .NET Framework is on Assess
44
its last release, Microsoft have reinforced 39.93
Anthos
92
the view that .NET Core is the future of 49 45 40. Apache Pulsar
98
.NET. Microsoft has done a lot of work to 47 41. Cosmos
46 48 91
make .NET Core container friendly. Most of 42. Google BigQuery ML
our .NET Core–based projects target Linux 90 43. JupyterLab
and are often deployed as containers. The 44. Marquez
upcoming .NET 5 release looks promising, 45. Matomo
and we’re looking forward to it. 46. MeiliSearch
47. Stratos
48. Trillian
Istio
Adopt Hold
49. Node overload
If you’re building and operating a
scaled microservices architecture and
have embraced Kubernetes, adopting resiliency. Its user experience has been Anka
service mesh to manage all cross-cutting improved in its latest releases, because Trial
aspects of running the architecture of its ease of installation and control
is a default position. Among various panel architecture. Istio has lowered Anka is a set of tools to create, manage,
implementations of service mesh, Istio the bar for implementing large-scale distribute, build and test macOS
has gained majority adoption. It has microservices with operational quality for reproducible virtual environments for iOS
a rich feature set, including service many of our clients, while admitting that and macOS. It brings Docker-like experience
discovery, traffic management, service- operating your own Istio and Kubernetes to macOS environments: instant start, CLI
to-service and origin-to-service security, instances requires adequate knowledge to manage virtual machines and registry
observability (including telemetry and and internal resources which is not for the to version and tag virtual machines for
distributed tracing), rolling releases and fainthearted. distribution. We’ve used Anka to build a

15 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Platforms macOS private cloud for a client. This tool is
worth considering when virtualizing iOS and
the platform streamlines the text that
needs translation into an online workflow.
that’s supported by Google’s underlying
scalable infrastructure. We particularly like
macOS environments. We like that Crowdin nudges the teams to Firebase App Distribution, which makes
continuously and incrementally incorporate it easy to publish test versions of an app
translations rather than managing them in via a CD pipeline, and Firebase Remote
Although this technology Argo CD large batches toward the end. Config, which allows configuration changes
is not new, it is now Trial to be dynamically pushed to apps without
coming into its own with needing to republish them.
the increasing use of Without making a judgment of the GitOps eBPF
microservices deployed as technique, we’d like to talk about Argo Trial
orchestrated containers. CD within the scope of deploying and Hot Chocolate
monitoring applications in Kubernetes For several years now, the Linux kernel Trial
environments. Based on its ability to has included the extended Berkeley
(eBPF)
automate the deployment of the desired Packet Filter (eBPF) virtual machine and The GraphQL ecosystem and community
application state in the specified target provided the ability to attach eBPF filters keep growing. Hot Chocolate is a GraphQL
environments in Kubernetes and our good to particular sockets. But extended BPF server for .NET (Core and Classic). It lets
The clear separation of experience with troubleshooting failed goes far beyond packet filtering and allows you build and host schemas and then serve
identity from the rest of deployments, verifying logs and monitoring custom scripts to be triggered at various queries against them using the same base
the OAuth2 framework deployment status, we recommend you give points within the kernel with very little components of GraphQL — data loader,
makes it easier to Argo CD a try. You can even see graphically overhead. Although this technology isn’t resolver, schema, operations and types. The
integrate Hydra with an what is going on in the cluster, how a new, it’s now coming into its own with the team behind Hot Chocolate has recently
existing authentication change is propagated and how pods are increasing use of microservices deployed as added schema stitching, which allows
created and destroyed in real time. orchestrated containers. Service-to-service for a single entry point to query across
ecosystem.
communications can be complex in these multiple schemas aggregated from different
systems, making it difficult to correlate locations. Despite the potential to misuse
(Hydra)
Crowdin latency or performance issues back to an this approach, our teams are happy with
Trial API call. We’re now seeing tools released Hot Chocolate — it’s well documented, and
with prewritten eBPF scripts for collecting we’re able to deliver value quickly to our
Most of the projects with multilingual and visualizing packet traffic or reporting on clients.
support start with development teams CPU utilization. With the rise of Kubernetes,
building features in one language and we’re seeing a new generation of security
managing the rest through offline enforcement and instrumentation based on Hydra
translation via emails and spreadsheets. eBPF scripts that help tame the complexity Trial
Although this simple setup works, things of a large microservices deployment.
can quickly get out of hand. You may have Not everyone needs a self-hosted OAuth2
to keep answering the same questions solution, but if you do, have a look at
for different language translators, sucking Firebase Hydra — a fully compliant open source
the energy out of the collaboration Trial OAuth2 server and OpenID connect
between translators, proofreaders and provider. Hydra has in-memory storage
the development team. Crowdin is one Google’s Firebase has undergone significant support for development and a relational
of a handful of platforms that help in evolution since we mentioned it as part database (PostgreSQL) for production
streamlining the localization workflow of of a serverless architecture in 2016. use cases. Hydra as such is stateless and
your project. With Crowdin the development Firebase is a comprehensive platform for easy to scale horizontally in platforms
team can continue building features, while building mobile and web apps in a way such as Kubernetes. Depending on your

TECHNOLOGY RADAR | 16
© ThoughtWorks, Inc. All Rights Reserved.
 Platforms
performance requirement, you may have solution for many of our clients. It has providing a high-level management and
to tune the number of database instances a superior architecture to scale storage, control plane on top of a set of open
while scaling Hydra instances. And compute, and services to load, unload and source technologies such as GKE, Service
because Hydra doesn’t provide any identity use data. It’s also very flexible: it supports Mesh and a Git-based Configuration
management solutions out of the box, you storage of structured, semi-structured and Management. It enables running
can integrate whatever flavor of identity unstructured data; provides a growing list portable workloads and other assets The OpenTelemetry
management you have with Hydra through of connectors for different access patterns on different hosting environments, project includes
a clean API. This clear separation of identity such as Spark for data science and SQL including Google Cloud and on-premises specification, libraries,
from the rest of the OAuth2 framework for analytics; and runs on multiple cloud hardware. Although other cloud providers agents and other
makes it easier to integrate Hydra with an providers. Our advice to many of our have comparative offerings, Anthos components needed to
existing authentication ecosystem. clients is to use managed services for their intends to go beyond a hybrid cloud to
capture telemetry from
utility technology such as big data storage; a portable cloud enabler using open
however, if the risk and regulations source components, but that is yet to services to better observe,
OpenTelemetry prohibit the use of managed services, be seen. We’re seeing a rising interest manage and debug them.
Trial then Snowflake is a good candidate for in Anthos. While Google’s approach in
companies with large volumes of data and managed hybrid cloud environments (OpenTelemetry)
OpenTelemetry is an open source heavy processing workloads. Although seems promising, it’s not a magic bullet
observability project that merges we’ve been successful using Snowflake in and requires changes in both existing
OpenTracing and OpenCensus. The our medium-sized engagements, we’ve cloud and on-premise assets. Our advice Anthos is Google’s answer
OpenTelemetry project includes yet to experience Snowflake in large for clients considering Anthos is to make to enable hybrid and
specification, libraries, agents, and other ecosystems where data need to be owned measured tradeoffs between selecting multicloud strategies.
components needed to capture telemetry across segments of the organization. services from the Google Cloud ecosystem
It enables you to run
from services to better observe, manage and other options, to maintain their right
level of neutrality and control. portable workloads
and debug them. It covers the three pillars
of observability — distributed tracing, Anthos on different hosting
metrics and logging (currently in beta) — Assess environments including
and its specification connects these three Apache Pulsar Google Cloud and on-
pieces through correlations; thus you We see a shift from accidental hybrid or Assess premise hardware.
can use metrics to pinpoint a problem, whole-of-estate cloud migration plans
locate the corresponding traces to to intentional and sophisticated hybrid, Apache Pulsar is an open source pub- (Anthos)
discover where the problem occured, and poly or portable cloud strategies, where sub messaging/streaming platform,
ultimately study the corresponding logs to organizations apply multidimensional competing in a similar space with Apache
find the exact root cause. OpenTelemetry principles to establish and execute their Kafka. It provides expected functionality
components can be connected to back-end cloud strategy: where to host their various — such as low-latency async and sync
observability systems such as Prometheus data and functional assets based on message delivery and scalable persistent
and Jaeger among others. Formation of risk, ability to control and performance storage of messages — as well as various
OpenTracing is a positive step toward the profiles; how to utilize their on-premise client libraries. What has excited us to
convergence of standardization and the infrastructure investments while reducing evaluate Pulsar is its ease of scalability,
simplification of tooling. the cost of operations; and how to take particularly in large organizations with
advantage of multiple cloud providers and multiple segments of users. Pulsar natively
their unique differentiated services without supports multitenancy, georeplication,
Snowflake creating complexity and friction for users role-based access control and segregation
Trial building and operating applications. of billing. We’re also looking to Pulsar to
solve the problem of a never-ending log of
Snowflake has proven to be a robust Anthos is Google’s answer to enable messages for our large-scale data systems
SaaS big data storage, warehouse or lake hybrid and multicloud strategies by where events are expected to persist

17 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Platforms
indefinitely and subscribers are able to BigQuery ML inverts this by bringing the information about a data ecosystem. It
start consuming messages retrospectively. model to the data. Google BigQuery is a represents a simple data model to capture
This is supported through a tiered storage data warehouse designed to serve large- metadata such as lineage, upstream and
model. Although Pulsar is a promising scale queries using SQL, for analytical use downstream data processing jobs and their
platform for large organizations, there cases. Google BigQuery ML extends this status, and a flexible set of tags to capture
BigQuery ML lowers is room for improvement. Its current function and its SQL interface to create, the attributes of data sets. It provides a
the bar for using ML installation requires administering train and evaluate machine learning simple RESTful API to manage the metadata
to make predictions ZooKeeper and BookKeeper among other models using its data sets; and eventually which eases the integration of Marquez to
and recommendations, pieces of technology. We hope that with its run model predictions to create new other tool sets within the data ecosystem.
particularly for quick growing adoption, users can soon count on BigQuery data sets. It supports a limited
wider community support. set of models out of the box, such as linear We’ve used Marquez as a starting point
explorations.
regression for forecasting or binary and and easily extended it to fit our needs such
multiclass regression for classification. It as enforcing security policies as well as
(Google BigQuery ML)
Cosmos also supports, with limited functionality, changes to its domain language. If you’re
Assess importing previously trained TensorFlow looking for a small and simple tool to
models. Although BigQuery ML and its SQL- bootstrap storage and visualization of your
The performance of blockchain technology based approach lower the bar for using data-processing jobs and data sets, Marquez
has been greatly improved since we initially machine learning to make predictions and is a good place to start.
assessed this area in the Radar. However, recommendations, particularly for quick
there’s still no single blockchain that could explorations, this comes with a difficult
achieve “internet-level” throughput. As trade-off: compromising on other aspects of Matomo
various blockchain platforms develop, we’re model training such as ethical bias testing, Assess
seeing new data and value silos. That’s explainability and continuous delivery for
why cross-chain tech has always been a machine learning. Matomo (formerly Piwik) is an open source
key topic in the blockchain community: web analytics platform that provides you
the future of blockchain may be a network with full control over your data. You can self-
of independent parallel blockchains. This JupyterLab host Matomo and secure your web analytics
is also the vision of Cosmos. Cosmos Assess data from third parties. Matomo also makes
releases Tendermint and CosmosSDK to it easy to integrate web analytics data with
let developers customize independent JupyterLab is the next-generation web- your in-house data platform and lets you
blockchains. These parallel blockchains based user interface for Project Jupyter. build usage models that are tailored to your
could exchange value through the Inter- If you’ve been using Jupyter Notebooks, needs.
Blockchain Communication (IBC) protocol JupyterLab is worth a try; it gives you
and Peg-Zones. Our teams have had great an interactive environment for Jupyter
experiences with CosmosSDK, and the notebooks, code and data. We see it as an MeiliSearch
IBC protocol is maturing. This architecture evolution of Jupyter Notebook: it provides a Assess
could solve blockchain interoperability and better experience by extending its original
scalability issues. capabilities of allowing code, visualization MeiliSearch is a fast, easy-to-use and easy-
and documentation to exist in one place. to-deploy text search engine. Over the
years Elasticsearch has become the popular
Google BigQuery ML choice for scalable text searches. However,
Assess Marquez if you don’t have the volume of data that
Assess warrants a distributed solution but still
Often training and predicting outcomes want to provide a fast typo-tolerant search
from machine learning models require Marquez is a relatively young open source engine, then we recommend assessing
code to take the data to the model. Google project for collecting and serving metadata MeiliSearch.

TECHNOLOGY RADAR | 18
© ThoughtWorks, Inc. All Rights Reserved.
Stratos
Assess
blockchain-based distributed ledgers. For
enterprise environments, however, where
popular, it was the first major framework to
embrace a nonblocking programming model Platforms
the cost of CPU-heavy consensus protocols which made it very efficient for IO-heavy
Ultraleap (previously Leap Motion) has is unwarranted, we recommend you give tasks. (We mentioned this in our write-up of
been a leader in the XR space for some Trillian a try. Node.js in 2012.) Due to its single-threaded
time, creating remarkable hand-tracking nature, Node.js was never a good choice MeiliSearch is a fast, easy-
hardware that allows a user’s hands to for compute-heavy workloads, though, and to-use and easy-to-deploy
make the leap into virtual reality. Stratos is Node overload now that capable nonblocking frameworks text search engine. It’s
Ultraleap’s underlying haptics, sensors and Hold also exist on other platforms — some with ideal if you don’t have
software platform, and it can use targeted elegant, modern APIs — performance is no the volume of data that
ultrasound to create haptic feedback in mid- Technologies, especially wildly popular ones, longer a reason to choose Node.js.
warrants a distributed
air. A use case is responding to a driver’s have a tendency to be overused. What we’re
solution but still want
hand gesture to change the air conditioning seeing at the moment is Node overload, a
in the car and providing haptic feedback tendency to use Node.js indiscriminately to provide a fast typo-
as part of the interface. We’re excited to or for the wrong reasons. Among these, tolerant search engine.
see this technology and what creative two stand out in our opinion. Firstly, we
technologists might do to incorporate it into frequently hear that Node should be used (MeiliSearch)
their use cases. so that all programming can be done in
one programming language. Our view
remains that polyglot programming is a Stratos is the underlying
Trillian better approach, and this still goes both haptics, sensors and
Assess ways. Secondly, we often hear teams cite software platform that
performance as a reason to choose Node.js.
powers XR pioneer
Trillian is a cryptographically verifiable, Although there are myriads of more or less
sensible benchmarks, this perception is Ultraleap (previously Leap
centralized data store. For trustless,
decentralized environments, you can use rooted in history. When Node.js became Motion).
(Stratos)

19 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
TECHNOLOGY RADAR Vol. 22

Tools
 Tools
Adopt
50. Cypress
27 51. Figma

26 Trial
52. Dojo
Cypress 22
67
68 53. DVC
Adopt 54. Experiment tracking tools
21 69 for machine learning
70
Cypress is still 25
a favorite among our teams 55. Goss
where developers manage end-to-end 56. Jaeger
tests themselves, as part of 20
a healthy test 52 57. k9s
54 71
pyramid, of course. We decided to call it out 15 58. kind
again in this Radar because recent versions 16 53 55 72 59. mkcert
13
24
of Cypress 19 support for Firefox,
have added 14 58
60. MURAL
and we strongly suggest testing on multiple
12 57 61. Open Policy Agent (OPA)
browsers. The dominance of Chrome and 73 62. Optimal Workshop
56 59
Chromium-based
18 browsers has led to a 63. Phrase
worrying trend of teams seemingly
10
11
only 74 64. ScoutSuite
testing with Chrome which can lead to nasty 6 60 62 65. Visual regression testing
23 61
surprises. 5 tools
9 63 75
66. Visual Studio Live Share
17 50
8 3 64
Figma 4 51 65 76 Assess
1 67. Apache Superset
Adopt 7 2 66
68. AsyncAPI
Hold FigmaAssess
has demonstratedTrial
to be the go-to Adopt Adopt Trial Assess Hold 69. ConfigCat
tool for collaborative design, not only for 70. Gitpod
designers but for multidisciplinary teams 28 89 71. Gloo
39 30 79 72. Lens
too; it allows developers and other roles
to view and comment on designs through 88
97 73. Manifold
87
the browser without 31the desktop version. 78 74. Sizzy
Dojo Docker images. Several of our teams use
75. Snowpack
Compared40 to its competitors (e.g., Invision 77 Dojo to streamline developing, testing
Trial
or Sketch) which have you use 32 more than 29 96
and building code from local development
76. tfsec
41 86
one tool for versioning, collaborating
33 and A few years ago, Docker — and containers in through production pipelines.
design sharing, Figma puts together all 34of 85 Hold
general — radically changed how we think
these features in one tool that makes it 83
about packaging, deploying and running84 our 95
42
easier for our teams to discover new ideas
35 applications. But despite
81
this improvement DVC
together. Our teams find Figma very useful, Trial
43 37in production, developers still spend a lot of 94
especially in remote and distributed design time setting up development environments
36 80 82
work enablement and facilitation. In addition 38
and regularly run into “but it works on my In 2018 we mentioned DVC in conjunction
to its real-time design and collaboration machine” style problems. Dojo aims to fix with the versioning data for reproducible
44
capabilities, Figma also offers an API that this by creating standard development
93
analytics. Since then it has become a
helps to improve49the DesignOps 92
45process. environments, versioned and released as favorite98tool for managing experiments in
46 47 48 91

90

21 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Tools machine learning (ML) projects. Since it’s
based on Git, DVC is a familiar environment
features, you may want to consider it when
its features meet your needs, especially
tool such as the AWS web console can be
a useful addition. It allows us to explore
for software developers to bring their since it comes as a small, self-contained all kinds of resources in an ad-hoc fashion
engineering practices to ML practice. binary (rather than requiring a Ruby without having to remember every single
Because it versions the code that processes environment). A common anti-pattern with obscure command. Using an interactive
Jaeger is an open source data along with the data itself and tracks using tools such as Goss is double-entry tool to make manual modifications on
distributed tracing system. stages in a pipeline, it helps bring order to bookkeeping, where each change in the the fly is still a questionable practice,
We’ve used it successfully the modeling activities without interrupting actual infrastructure as code files requires a though. For Kubernetes we now have k9s,
with Istio and Envoy on the analysts’ flow. corresponding change in the test assertions. which provides an interactive interface for
Kubernetes, and like its UI. Such tests are maintenance heavy and basically everything that kubectl can do.
because of the close correspondence And to boot, it’s not a web application but
(Jaeger) Experiment tracking tools for between code and test, failures mostly occur runs inside a terminal window, evoking
machine learning when an engineer updates one side and fond memories of Midnight Commander
Trial forgets the other. And these tests rarely for some of us.
catch genuine problems.
k9s provides an
The day-to-day work of machine
interactive interface kind
learning often boils down to a series
for basically everything of experiments in selecting a modeling Jaeger Trial
that kubectl can do. And approach and the network topology, Trial
to boot, it’s not a web training data and optimizing or tweaking kind is a tool for running local Kubernetes
application but runs the model. Data scientists must use Jaeger is an open source distributed clusters using Docker container nodes.
inside a terminal window. experience and intuition to hypothesize tracing system. Similar to Zipkin, it’s been With kubetest integration, kind makes
changes and then measure the impact inspired by the Google Dapper paper and it easy to do end-to-end testing on
(k9s) those changes have on the overall complies with OpenTelemetry. We’ve used Kubernetes. We’ve used kind to create
performance of the model. As this practice Jaeger successfully with Istio and Envoy on ephemeral Kubernetes clusters to test
has matured, our teams have found an Kubernetes and like its UI. Jaeger exposes Kubernetes resources such as Operators
increasing need for experiment tracking tracing metrics in the Prometheus format and Custom Resource Definitions (CRDs) in
tools for machine learning. These tools so they can be made available to other our CI pipelines.
help investigators keep track of the tools. However, a new generation of tools
experiments and work through them such as Honeycomb integrates traces and
methodically. Although no clear winner metrics into a single observability stream mkcert
has emerged, tools such as MLflow and for simpler aggregate analysis. Jaeger Trial
platforms such as Comet or Neptune have joined CNCF in 2017 and has recently been
introduced rigor and repeatability into the elevated to CNCF’s highest level of maturity, mkcert is a convenient tool for creating
entire machine learning workflow. indicating its widespread deployment into locally trusted development certificates.
production systems. Using certificates from real certificate
authorities (CAs) for local development can
Goss be challenging if not impossible (for hosts
Trial k9s such as example.test, localhost or 127.0.0.1).
Trial In such situations self-signed certificates may
We mentioned Goss, a tool for provisioning be your only option. mkcert lets you generate
testing, in passing in previous Radars, for We continue to be ardent supporters of self-signed certificates and installs the local
example, when describing the technique infrastructure as code, and we continue to CA in the system root store. For anything
of TDD’ing containers. Although Goss isn’t believe that a robust monitoring solution other than local development and testing, we
always an alternative to Serverspec, simply is a prerequisite for operating distributed strongly recommend using certificates from
because it doesn’t offer the same amount of applications. Sometimes an interactive real CAs to avoid trust issues.

TECHNOLOGY RADAR | 22
© ThoughtWorks, Inc. All Rights Reserved.
MURAL Optimal Workshop
Tools
GCP and other cloud providers. It works
Trial Trial by automatically aggregating configuration
data for an environment and applying rules
MURAL describes itself as a “digital UX research demands data collection and to audit the environment. We’ve found this
workspace for visual collaboration” and analysis to make better decisions about very useful across projects for doing point-
allows teams to interact with a shared the products we need to build. Our teams in-time security assessments. OPA provides a uniform
workspace based on a whiteboard/sticky find Optimal Workshop useful because it framework and language
notes metaphor. Its features include makes it easy to validate prototypes and for declaring, enforcing
voting, commenting, notes and “follow the configure tests for data collection and Visual regression testing tools and controlling policies
presenter.” We particularly like the template thus make better decisions. Features such Trial for various components of
feature that allows a facilitator to design as first-click, card sorting, or a heatmap a cloud-native solution.
and then reuse guided sessions with a team. of user interaction help to both validate Since we first mentioned visual regression
Each of the major collaboration suites have prototypes and improve website navigation testing tools in 2014, the use of the
(Open Policy Agent (OPA))
a tool in this space (for example, Google and information display. It’s an ideal tool technique has spread and the tools
Jamboard and Microsoft Whiteboard) and for distributed teams since it allows them to landscape has evolved. BackstopJS remains
these are worth investigating, but we’ve conduct remote research. an excellent choice with new features being
found MURAL to be slick, effective and added regularly, including support for Our teams find Optimal
flexible. running inside Docker containers. Loki was Workshop useful because
Phrase featured in our previous Radar. Applitools, it makes it easy to validate
Trial CrossBrowserTesting and Percy are SaaS prototypes and configure
Open Policy Agent (OPA) solutions. Another notable mention is tests for data collection
Trial As mentioned in our description of Crowdin, Resemble.js, an image diffing library. and thus make better
you now have a choice of platforms to Although most teams use it indirectly as
decisions.
Open Policy Agent (OPA) has rapidly manage the translation of a product into part of BackstopJS, some of our teams
become a favorable component of many multiple languages instead of emailing large have been using it to analyze and compare
(Optimal Workshop)
distributed cloud-native solutions that we spreadsheets. Our teams report positive images of web pages directly. In general,
build for our clients. OPA provides a uniform experiences with Phrase, emphasizing our experience shows that visual regression
framework and language for declaring, that it’s easy to use for all key user groups. tools are less useful in the early stages
enforcing and controlling policies for various Translators use a convenient browser- when the interface goes through significant
components of a cloud-native solution. It’s based UI. Managers can add new fields and changes, but they certainly prove their
a great example of a tool that implements synchronize translations with other teams worth as the product matures and the
security policy as code. We’ve had a smooth in the same UI. Developers can access interface stabilizes.
experience using OPA in multiple scenarios, Phrase locally and from a build pipeline. A
including deploying resources to K8s feature that deserves a specific mention is
clusters, enforcing access control across the ability to apply versioning to translations Visual Studio Live Share
services in a service mesh and fine-grained through tags, which makes it possible to Trial
security controls as code for accessing compare the look of different translations
application resources. A recent commercial inside the actual product. Visual Studio Live Share is a suite of
offering, Styra’s Declarative Authorization extensions for Visual Studio Code and Visual
Service (DAS), eases the adoption of OPA Studio. At a time when teams are searching
for enterprises by adding a management ScoutSuite for good remote collaboration options,
tool, or control plane, to OPA for K8s with Trial we want to call attention to the excellent
a prebuilt policy library, impact analysis of tooling here. Live Share provides a good,
the policies and logging capabilities. We look ScoutSuite is an expanded and updated low-latency remote-pairing experience, and
forward to maturity and extension of OPA tool based on Scout2 (featured in the requires significantly less bandwidth than
beyond operational services to (big) data- Radar in 2018) that provides security the brute-force approach of sharing your
centric solutions. posture assessment across AWS, Azure, entire desktop. Importantly, developers can

23 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Tools Gloo
work with their preferred configuration, specification, inspired by the OpenAPI
extensions and key mappings during a specification, describes and documents Assess
pairing session. In addition to real-time event-driven APIs in a machine-readable
collaboration for editing and debugging format. It’s protocol agnostic, so it can With the increasing adoption of Kubernetes
code, Live Share allows voice calls and be used for APIs that work over many and service mesh, API gateways have been
AsyncAPI is an open sharing terminals and servers. protocols, including MQTT, WebSockets, experiencing an existential crisis in cloud-
source initiative to create and Kafka. We’re eager to see the ongoing native distributed systems. After all, many
a much needed event- improvements of AsyncAPI and further of their capabilities (such as traffic control,
driven and asynchronous Apache Superset maturity of its tooling ecosystem. security, routing and observability) are now
API standardization and Assess provided by the cluster’s ingress controller
development tooling. and mesh gateway. Gloo is a lightweight
Apache Superset is a great business ConfigCat API gateway that embraces this change;
intelligence (BI) tool for data exploration Assess it uses Envoy as its gateway technology,
(AsyncAPI)
and visualization to work with large data while providing added value such as a
lake and data warehouse setups. It works, If you’re looking for a service to support cohesive view of the APIs to the external
for example, with Presto, Amazon Athena dynamic feature toggles (and bear in mind users and applications. It also provides
ConfigCat supports and Amazon Redshift and can be nicely that simple feature toggles work well too), an administrative interface for controlling
simple feature toggles, integrated with enterprise authentication. check out ConfigCat. We’d describe it as “like Envoy gateways and runs and integrates
user segmentation, and Moreover, you don’t have to be a data LaunchDarkly but cheaper and a bit less with multiple service mesh implementations
A/B testing and has a engineer to use it; it’s meant to benefit fancy” and find that it does most of what such as Linkerd, Istio and AWS App Mesh.
generous free tier for low- all engineers exploring data in their we need. ConfigCat supports simple feature While its open source implementation
volume use cases or those everyday work. It’s worth pointing out that toggles, user segmentation, and A/B testing provides the basic capabilities expected
Apache Superset is currently undergoing and has a generous free tier for low-volume from an API gateway, its enterprise edition
just starting out.
incubation at the Apache Software use cases or those just starting out. has a more mature set of security controls
Foundation (ASF), meaning it’s not yet fully such as API key management or integration
(ConfigCat)
endorsed by ASF. with OPA. Gloo is a promising lightweight
Gitpod API gateway that plays well with the
Assess ecosystem of cloud-native technology and
AsyncAPI architecture, while avoiding the API gateway
Assess You can build most software following trap of enabling business logic to glue APIs
a simple two-step process: check out a for the end user.
Open standards are one of the foundational repository, and then run a single build
pillars of building distributed systems. For script. The process of setting up a full coding
example, the OpenAPI (formerly Swagger) environment can still be cumbersome, Lens
specification, as an industry standard to though. Gitpod addresses this by providing Assess
define RESTful APIs, has been instrumental cloud-based, “ready-to-code” environments
to the success of distributed architectures for Github or GitLab repositories. It offers One of the strengths of Kubernetes is
such as microservices. It has enabled a an IDE based on Visual Studio Code that its flexibility and range of configuration
proliferation of tooling to support building, runs inside the web browser. By default, possibilities along with the API-driven,
testing and monitoring RESTful APIs. these environments are launched on the programmable configuration mechanisms
However, such standardizations have been Google Cloud Platform, although you can and command-line visibility and control
largely missing in distributed systems for also deploy on-premise solutions. We see using manifest files. However, that strength
event-driven APIs. the immediate appeal, especially for open can also be a weakness: when deployments
source software where this approach can are complex or when managing multiple
AsyncAPI is an open source initiative lower the bar for casual contributors. clusters, it can be difficult to get a clear
to create a much needed event-driven However, it remains to be seen how picture of the overall status through
and asynchronous API standardization viable this approach will be in corporate command-line arguments and manifests
and development tooling. The AsyncAPI environments. alone. Lens attempts to solve this problem

TECHNOLOGY RADAR | 24
© ThoughtWorks, Inc. All Rights Reserved.
 Sizzy
Tools
with an integrated environment for viewing feedback cycle during development because
the current state of the cluster and its Assess changes become available in the browser
workloads, visualizing cluster metrics almost immediately. For this magic to work,
and changing configurations through Building web applications that look just as Snowpack transforms the dependencies
an embedded text editor. Rather than a intended on a large number of devices and in node_modules into single JavaScript
simple point-and-click interface, Lens brings screen sizes can be cumbersome. Sizzy is a files in a new web_modules directory, Lens is one of several
together the tools an administrator would SaaS solution that shows many viewports from where they can be imported as an different approaches that
run from the command line into a single, in a single browser window. The application ECMAScript module (ESM). For IE11 and are trying to tame the
navigable interface. This tool is one of is rendered in all viewports simultaneously other browsers that don’t support ESM, a complexity of Kubernetes
several approaches that are trying to tame and interactions with the application are workaround is available. Unfortunately, management.
the complexity of Kubernetes management. also synched across the viewports. In our because no browser today can import CSS
We’ve yet to see a clear winner in this space, experience interacting with an application in from JavaScript, using CSS modules is not
(Lens)
but Lens strikes an interesting balance this way can make it easier to spot potential straightforward.
between a graphical UI and command-line– issues earlier, before a visual regression
only tools. testing tool flags the issue in the build
tfsec tfsec is a static analysis
pipeline. We should mention, though, that
some of our developers who tried Sizzy for Assess tool that helps to scan
Manifold a while did, on balance, prefer to work with Terraform templates and
Assess the tooling provided by Chrome. Security is everyone’s concern and capturing find any potential security
risks early is always better than facing issues.
Manifold is a model-agnostic visual problems later on. In the infrastructure as
debugger for machine learning (ML). Model Snowpack code space, where Terraform is an obvious (tfsec)
developers spend a significant amount of Assess choice to manage cloud environments,
time on iterating and improving an existing we now also have tfsec, which is a static
model rather than creating a new one. Snowpack is an interesting new entrant analysis tool that helps to scan Terraform
By shifting the focus from model space in the field of JavaScript build tools. The templates and find any potential security
to data space, Manifold supplements key improvement over other solutions is issues. It comes with preset rules for
the existing performance metrics with a that Snowpack makes it possible to build different cloud providers including AWS
visual characteristics of the data set that applications with modern frameworks such and Azure. We always like tools that help
influences the model performance. We think as React.js, Vue.js, and Angular without to mitigate security risks, and tfsec not only
Manifold will be a useful tool to assess in the the need for a bundler. Cutting out the excels in identifying security risks, it’s also
ML ecosystem. bundling step dramatically improves the easy to install and use.

25 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
TECHNOLOGY RADAR Vol. 22

Languages &
Frameworks
 23 61
5
9 63 75

17 50
8 3 64
4 51 65 76
1
7 2 66

Languages &
Hold Assess Trial Adopt Adopt Trial Assess Hold
Adopt
77. React Hooks
78. React Testing Library

Frameworks
28 89
39 30 79 79. Vue.js
97
88
31 78 87 Trial
40
80. CSS-in-JS
77
32 29 96 81. Exposed
41 86 82. GraphQL Inspector
33
34 83. Karate
React Hooks 85
84. Koin
83
Adopt 84 95 85. NestJS
42
35 81 86. PyTorch
React Hooks have introduced a 43 new approach 37 94 87. Rust
to managing stateful logic; given React 36 80 82 88. Sarama
components have always been closer to 38
89. SwiftUI
functions than classes, Hooks have44embraced
this and brought state to the functions, 93
92 Assess
instead of taking function49as methods to45the 98
90. Clinic.js Bubbleprof
state with classes. Based on our experience, 46 47 48 91 91. Deequ
Hooks improve reuse of functionality among 92. ERNIE
components and code readability. Given 90 93. MediaPipe
Hooks’ testability improvements, using React 94. Tailwind CSS
Test Renderer and React Testing Library, 95. Tamer
and their growing community support, we 96. Wire
consider them our approach of choice. 97. XState

Hold
React Testing Library 98. Enzyme
Adopt
Library which React Testing Library is part segregation of directives and components
The JavaScript world moves pretty fast, of and which provides a whole family of (one file per component idiom) and its
and as we gain more experience using a libraries for Angular and Vue.js, for example. simpler state management have made it a
framework our recommendations change. compelling option among others.
The React Testing Library is a good example
of a framework that with deeper usage has
eclipsed the alternatives to become the
Vue.js
sensible default when testing React-based
Adopt CSS-in-JS
Trial
frontends. Our teams like the fact that tests
Vue.js has become one of the successfully
written with this framework are less brittle
applied, loved and trusted frontend Since we first mentioned CSS-in-JS as
than with alternative frameworks such as
JavaScript frameworks among our an emerging technique in 2017, it has
Enzyme, because you’re encouraged to
community. Although there are other, well- become much more popular, a trend we
test component relationships individually
adopted alternatives, such as React.js, the also see in our work. With some solid
as opposed to testing all implementation
simplicity of Vue.js in API design, its clear production experience under our belts,
details. This mindset is brought by Testing

27 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Languages & we can now recommend CSS-in-JS as a
technique to trial. A good starting point is
We’ve cautioned against the use of GraphQL
in the past, and we’re happy to see some
constructors or by mimicking Kotlin’s lazy
initialization so that objects are injected

Frameworks the styled components framework, which

we mentioned in our previous Radar. Next
improvements in tooling around GraphQL
since. Most of our teams continue to
only when needed. This is in contrast to
the statically compiled Dagger injection
to all the positives, though, there usually use GraphQL for server-side resource framework for Android. Our developers like
is a downside when using CSS-in-JS: the aggregation, and by integrating GraphQL the lightweight nature of this framework
calculation of styles at runtime can cause Inspector in their CI pipelines, we’ve been and its built-in testability.
a noticeable lag for end users. With Linaria able to catch potential breaking changes in
Koin is a Kotlin framework
we’re now seeing a new class of frameworks the GraphQL schema.
that handles one of the that were created with this issue in mind. NestJS
routine problems in Linaria employs a number of techniques Trial
software development, to shift most of the performance overhead Karate
dependency injection. to build time. Alas, this does come with its Trial The growth in popularity of Node.js
own set of trade-offs, most notably a lack of and trends such as Node overload have
(Koin) dynamic style support in IE11. Given our experience that tests are the only led to the application of Node.js for
API specifications that really matter, we’re developing business applications. We
always on the lookout for new tools that often see problems, such as scalability and
NestJS is a TypeScript-first Exposed might help with testing. Karate is an API maintainability, with large JavaScript-based
framework that makes the Trial testing framework whose unique feature applications. NestJS is a TypeScript-first
is that tests are written directly in Gherkin framework that makes the development
development of NodeJS
Through their extended use of Kotlin, our without relying on a general-purpose of Node.js applications safer and less error
applications safer and less development teams have gained experience programming language to implement test prone. NestJS is opinionated and comes with
error prone. with more frameworks designed specifically behavior. It’s a domain-specific language for SOLID principles and an Angular-inspired
for Kotlin rather than using Java frameworks describing HTTP-based API tests. Our teams architecture out of the box. When building
(NestJS) with Kotlin. Although it’s been around for a like the readable specification that they get Node.js microservices, NestJS is one of the
while, Exposed has caught our attention as a with this tool and recommend to keep tests frameworks that our teams commonly
lightweight object-relational mapper (ORM). with Karate in the upper levels of the testing use to empower developers to create
Exposed has two flavors of database access: pyramid and not overload its use by making testable, scalable, loosely coupled and easily
a typesafe internal DSL wrapping SQL and very detailed assertions. maintainable applications.
an implementation of the data access object
(DAO) pattern. It supports features expected
from a mature ORM such as handling of Koin PyTorch
many-to-many references, eager loading, Trial Trial
and support for joins across entities. We
also like that the implementation works As Kotlin is used increasingly for both Our teams have continued to use and
without proxies and doesn’t rely on mobile and server-side development, the appreciate the PyTorch machine learning
reflection, which is certainly beneficial to associated ecosystem continues to evolve. framework, and several teams prefer
performance. Koin is a Kotlin framework that handles PyTorch over TensorFlow. PyTorch exposes
one of the routine problems in software the inner workings of ML that TensorFlow
development: dependency injection. hides, making it easier to debug, and
GraphQL Inspector Although you can choose from a variety contains constructs that programmers
Trial of dependency injection frameworks for are familiar with such as loops and
Kotlin, our teams have come to prefer actions. Recent releases have improved
GraphQL Inspector lets you compare the simplicity of Koin. Koin avoids using performance of PyTorch, and we’ve been
changes between two GraphQL schemas. annotations and injects either through using it successfully in production projects.

TECHNOLOGY RADAR | 28
© ThoughtWorks, Inc. All Rights Reserved.
Rust
Trial
for a much better developer experience.
The SwiftUI framework also draws
tools in this space. They settled on Deequ,
a library for writing tests that resemble Languages &
Rust is continuously gaining in popularity.
inspiration from the React.js world that
has dominated web development in recent
unit tests for data sets. Deequ is built on
top of Apache Spark, and even though it’s Frameworks
We’ve had heated discussions about which years. Immutable values in view models and published by AWS Labs it can be used in
is better, Rust or C++/Go, without a clear an asynchronous update mechanism make environments other than AWS.
winner. However, we’re glad to see Rust for a unified reactive programming model.
has improved significantly, with more This gives developers an entirely native Deequ is a library for
built-in APIs being added and stabilized, alternative to similar reactive frameworks ERNIE writing tests that resemble
including advanced async support, since such as React Native or Flutter. SwiftUI Assess
definitely represents the future of Apple
unit tests for data sets
we mentioned it in our previous Radar. In
addition, Rust has also inspired the design UI development, and although new, it has In the previous edition of the Radar we
which can help when
of new languages. For example, the Move shown its benefits. We’ve been having great had BERT — which is a key milestone in the automating data quality
language on Libra borrows Rust’s way of experience with it — and its shallow learning NLP landscape. Last year, Baidu released checks between different
managing memory to manage resources, curve. It’s worth noting that you should ERNIE 2.0 (Enhanced Representation steps in a data pipeline.
ensuring that digital assets can never be know your customer’s use case before through kNowledge IntEgration) which
copied or implicitly discarded. jumping into using SwiftUI, given that it outperformed BERT on seven GLUE (Deequ)
doesn’t support iOS 12 or below. language understanding tasks and on all
nine of the Chinese NLP tasks. ERNIE, like
Sarama BERT, provides unsupervised pretrained ERNIE provides
Trial Clinic.js Bubbleprof language models, which can be fine-
unsupervised pretrained
Assess tuned by adding output layers to create
state-of-the-art models for a variety of
language models, which
Sarama is a Go client library for Apache
Kafka. If you’re developing your APIs in Go, With the aim of improving performance NLP tasks. ERNIE differs from traditional can be fine-tuned by
you’ll find Sarama quite easy to set up and in our code, profiling tools are very useful pretraining methods in that it is a continual adding output layers to
manage as it doesn’t depend on any native to identify bottlenecks or delays in code pretraining framework. Instead of training create state-of-the-art
libraries. Sarama has two types of APIs — which are hard to identify, especially with a small number of pretraining models for a variety of
a high-level API for easily producing and in asynchronous operations. Clinic.js objectives, it could constantly introduce NLP tasks.
consuming messages and a low-level API for Bubbleprof represents visually the async a large variety of pretraining tasks to
controlling bytes on the wire. operations in Node.js processes, drawing a help the model efficiently learn language (ERNIE)
map of delays in the application’s flow. We representations. We’re pretty excited about
like this tool because it helps developers to the advancements in NLP and are looking
SwiftUI easily identify and prioritize what to improve forward to experimenting with ERNIE on our
Trial in the code. projects.

Apple has taken a big step forward

with their new SwiftUI framework for Deequ MediaPipe
implementing user interfaces on the macOS Assess Assess
and iOS platforms. We like that SwiftUI
moves beyond the somewhat kludgy There are still some tool gaps when applying MediaPipe is a framework for building
relationship between Interface Builder and good software engineering practices in MultiModal (such as video, audio, time
Xcode and adopts a coherent, declarative data engineering. Attempting to automate series data, etc.), cross-platform (for
and code-centric approach. You can now data quality checks between different example, Android, iOS, Web, and edge
view your code and the resulting visual steps in a data pipeline, one of our teams devices) and applied ML pipelines. It
interface side by side in Xcode 11, making was surprised when they found only a few provides multiple capabilities, including face

29 | TECHNOLOGY RADAR
© ThoughtWorks, Inc. All Rights Reserved.
Languages & detection, hand tracking, gesture detection
and object detection. Although MediaPipe
JDBC source connector for Kafka.” Despite
being a relatively new framework, we’ve
visualizing them as state charts. It integrates
with the more popular reactive JavaScript

Frameworks is primarily deployed to mobile devices, it’s

started to show up in the browser thanks to
found Tamer to be more efficient than the
Kafka JDBC connector, especially when huge
frameworks (Vue.js, Ember.js, React.js and
RxJS) and is based on the W3C standard
WebAssembly and XNNPack ML Inference amounts of data are involved. for finite state machines. Another notable
Library. We’re exploring MediaPipe for some feature is the serialization of machine
AR use cases and like what we see so far. definitions. One thing that we’ve found
Tailwind CSS proposes Wire helpful when creating finite state machines
an interesting approach Assess in other contexts (particularly when writing
by providing lower-level Tailwind CSS game logic) is the ability to visualize states
Assess The Golang community has had its fair and their possible transitions; we like the
utility CSS classes to fact that it’s really easy to do this with
share of dependency injection skeptics,
create building blocks CSS tools and frameworks offer predesigned partly because they confused the pattern XState’s visualizer.
without opinionated components for fast results; after a with specific frameworks, and developers
styles and aiming for easy while, however, they can complicate with a system-programming background
customization. customization. Tailwind CSS proposes an naturally dislike runtime overhead caused Enzyme
interesting approach by providing lower- by reflection. Then along came Wire, a Hold
(Tailwind CSS) level utility CSS classes to create building compile-time dependency injection tool that
blocks without opinionated styles and can generate code and wire components We don’t always move deprecated tools
aiming for easy customization. The breadth together. Wire has no additional runtime to Hold in the Radar, but our teams feel
Wire is a compile-time of the low-level utilities allows you to overhead, and the static dependency graph strongly that Enzyme has been replaced for
avoid writing any classes or CSS on your is easier to reason about. Whether you unit testing React UI components by React
dependency injection tool
own which leads to a more maintainable handwrite your code or use frameworks, we Testing Library. Teams using Enzyme have
that can both generate codebase in the long term. It seems that recommend using dependency injection to found that its focus on testing component
code and wire components Tailwind CSS offers the right balance encourage modular and testable designs. internals leads to brittle, unmaintainable
together. between reusability and customization to tests.
create visual components.
(Wire) XState
Assess
Tamer
Assess We’ve featured several state management
libraries in the Radar before, but XState
If you need to ingest data from relational takes a slightly different approach. It’s a
databases into a Kafka topic, consider simple JavaScript and TypeScript framework
Tamer, which labels itself “a domesticated for creating finite state machines and

TECHNOLOGY RADAR | 30
© ThoughtWorks, Inc. All Rights Reserved.
We are a software consultancy and Want to stay up-to-date with all
community of passionate purpose-led Radar-related news and insights?
individuals, 7000+ people strong across
Follow us on your favorite social channel or
43 offices in 14 countries. Over our 25+
become a subscriber.
year history, we have helped our clients
solve complex business problems
where technology is the differentiator.
subscribe now
When the only constant is change, we
prepare you for the unpredictable.
thoughtworks.com/radar
#TWTechRadar