Jun 1st, 2018

DotScale 2018 (Yet Another Conference Blogpost πŸ™ˆ)

conference, dotscale, Paris, people

Once again the DotConferences team organised a really nice one-day one-track conference in Paris. Thanks a lot to them, to the speakers and the audience. DotScale 2018 was a very fine edition!

They managed to have very nice and experienced speakers, from diverse backgrounds (academic, industry, tech advocates) and with various subjects. It was my fifth DotConference since I live in Paris! Wow time flies by.

You will find below a few notes of the day if you want to have a general idea of what happened there.

If you prefer to watch the complete talks, videos should be made available from the DotConferences team.

Single day - Single Track

Talks

Istio, We have a problem!

David Gageot - @dgageot

Istio is a set of components on top of Kubernetes to build a service mesh.

The Setup

What am I trying to achieve? Web developer building a simple app. Quiz website to find which image matches the description. Based on micro-services.

I am going to deploy the app using a tool called skaffold (build, tag and deployment tool to watch local directory)

The app seems to work sometimes and sometimes it’s broken. Β―\_(ツ)_/Β―

5 micro services (yes it’s crazy for this app). Houston we got a problem.

What is happening? I don’t know…

If only we had distributed traces, if only we had visibility…

One solution: Istio “Service Mesh”

Showing the flow of a user request via:

Istio knows what is happening in my micro services world. You can thus show a graph showing the relationships between them with a graphviz visualisation. That’s cool and “for free”.

Istio captures all traffic between services. It knows which endpoints are called, it knows their versions, it knows the http response time, response codes etc… All captured via prometheus and displayed in Grafana.

Istio installed on your cluster shows the global status of your micro-services platform.

David showcasing the metrics from each services aggregated in one graph: some services are not always responding with a correct http response code.

David Showcasing the tracing system via Zipkin. We can see that some requests have ugly traces (25 requests to a single service in loop it seems)

There’s a bug in that service! Let’s Fix It! Yeh and ship it!

Solution seems good for one user but seems broken for others.. When you fix something you always brake something else…! right?

David showcasing again the Grafana dashboards with the TWO versions of the code to see if it seems to fix the problem.

Let’s move all our users now that we see that it works!! Maybe not.

With istio you can move only a percentage of your traffic.

Sum up

Q/A

When will it be production ready?

Not yet, but V0.8 was out this morning! Healthier version ever

Other use cases

Adds security to your system (adds TLS layer in front of every micro-service). Another nice feature: Hybrid setup on-prem + cloud.

What are you working on personally right now?

Working on Skaffold to make developers happy on Kubernetes.

When will that be GA?

It’s open source halo, don’t know yet :p

↑ Back to the list of talks

Data Cartography: scaling the data landscape

Preeti Vaidya

Most important part of your app/solutions trying to build. Data is the key element of the things we build right? With data stores, cloud computing, data visualisation… It’s becoming really difficult to decide what to choose? What data storage should I choose? What data visualisation tool?

This talk should help you to choose!

Data is dynamic

That’s because human activity are dynamic!

How did I arrive here? NY β†’ Amsterdam β†’ Budapest ..... β†’ Paris

Spoke-hub model routing. Planing routes used in aviation. It’s also used in Healthcare. Also used in content marketing.

Trying to map human events where people are doing certain actions.

What is a data model?

It’s just an abstraction of your data. In order to fit in any new data in that data model.

E.g. with openflights.org How can we store this data?

First thing that comes to our mind: let’s use a relational model for this. Why do we need to go beyond this? BUT now you have a K/V store, a doc store, a graph store…

What DB do we choose?

Based on the e.g. of openflights.org if we try to visualise the data we can see some “hubs” appear. For example a certain city is routing lots of plane journeys. So what can be used is a graph database. Node which is our hub, edges which are our spokes. Possibly we can input our data in that data model. Graph first. Joins on this won’t be so computationally expensive because it’s designed to be easy to do.

E.g. debian packages as graph databases.

Two types of graph databases: native and non native.

Data apps are dynamic too

If we decided to go today to another place all of us together. Let’s see if a graph db would help us achieve that.

Our problem is now completely different: we need to find a consensus between all of us… Which will not be solved by a graph db. β†’ Byzantine generals problem. We want to agree on which place we would go.

A data model which solves this effectively is a Hashgraph data model. The hashgraph would propagate the information via neighbours. Eventually we would know which city gets more votes. Might be a good solution but then we would need to buy the tickets right?

So how could we do that? Maybe using a blockchain? Moving from our application to our operation. The operation that we are doing with our data.

Data operations are dynamic!

Top-down analysis

What data operation we want to do first. Then the application. Then the data model and finally the data store!

Sum-up

If we analyse first we would have the best data store for our solution. Don’t choose your data store first, think about your data usage before!

Q/A

How to put in practise this?

Taking relational data and put them together in one place. Think about the operations: enable people to do data analysis for certain products.

Book coming up? Tell us about it please :)

Bringing agility to data products. Data is a competitive advantage. It becomes important to do quicker analysis from your data. Move away from application design, to data product design.

Release of the book?

Soon hopefully :)

Thanks!

↑ Back to the list of talks

Distributions and package mngt in the containers era

Lucas Nussbaum - @lucasnussbaum

University of Loraine, Debian project leader

Distributions are not hype anymore.

FOSDEM 2017: 4 dev rooms for distributions

FOSDEM 2018: only 1!

Hype?

Everybody talking about it nobody using it VS everybody using it but nobody talking about it

Looking at failures of Distributions

How to distribute Linux Kernel, Gnome, GNU projects, Haproxy to end users?

Success: Universal package format

Most important thing about free open source software!

That’s thanks to a complex tool chain.

Debian leverage the language package managers (gems, pypy, npm, hackage…whatever). We can easily convert custom lang packages into debian packages. (wiki.debian.org/AutomaticPackagingTools)

Success: Consistent quality

Many QA tools.

Success: Distributions have ecosystem

Derivative of distributions: Ubuntu. Derivative distributions also has derivative distributions: Linux Mint

Failures: lack of collaboration

Problems/Issues should flow “back” to upstream projects. Due to derivatives and complex distribution graph sometimes it doesn’t come back all the way to the upstream packages.

Unclear status of distributions’ bug tracking systems.

Duplication info between all bug trackers between distributions and upstream packages.

In Debian issue tracking system bugs are not only “open/close” they are known as “fixed or found”. (For Multiple versions packages per distribution version)

Launchpad addresses the same Debian package management within Ubuntu packaging.

Should we have a federation of bug trackers?

Failures: usually not the software you want

Not the right version Not the right package

E.g. with containers

Solution: one big bundle

E.g. for ruby with bundler+gem2deb

Single big deb package for the two top layers (app + app dependencies). Already done by bundler or virtualenv for example. You can go further to bundle this in a deb package for example. gem2deb mygem!

E.g.2 auto.debian.net

Design a service that integrates all the tools via QA. To build everything automatically.

Q/A

Tell us more about the Debian project

Contributors all around the world. Very flat organisation. Some people says it’s an anarchist project. Everything is very flat :) One thing that makes it work: the debian constitution and policy documents + a social contract.

It’s not only “make the world a better place” :)

Relationship with Canonical and Ubuntu?

From the outside, people tends to think there’s lots of disagreements. But it’s not so much the case.

As we have priority for our users (it’s in the Debian policy) that’s the most important for us.

Most of Debian developers going to canonical stayed Debian contributors.

Of course canonical has their agenda. But they push back their fixes to us most of the time.

Advises for Debian contribution? How to get involved

Do it, it’s great! Packaging tutorial mentioned in my talk there’s a list of things to get started. You can join a team to help packaging. E.g. go in the ruby team.

↑ Back to the list of talks

Monitoring infra at scale

Damien Tournoud - @DamZ - CTO at Platform.sh

Monitoring infra that are highly dynamic.

Understanding, making sense of infrastructures. The more dynamic, the hardest time you’ll have to understand it (especially under pressure)

A step back

Platform.sh we unify infra to make it “simpler”. We are running applications for our clients. We rely on Azure, GCP, AWS… All hosts are divided into containers (we use big host with 500+ containers)

From the perspective of our clients: we deploy environments. We show a “logical view” as opposed to the “physical view” of our hosts.

Started small (launched in 2014). Now 15 regions. 8 cloud providers.

First step

Making a model of the components that we have.

Physical view, with region with zone. Logical view, services, envs, project.

in the middle we have regional services: impact on multiple customers or regionally located.

All this links Hosts with Instances.

Second step

Pull vs Push for monitoring?

For us Push made more sense.

Relationship aware push: Parent and child components. This helps to not keep track of “dead” components.

Alerts from metrics? In practice it’s really hard. It’s hard to design alerts that are useful (low level of false positive/false negatives).

For us we wanted the model to generate the alerting strategy. The model encapsulates our strategy.

Our end goal: figuring out how to feed the UI and waking up people at night (via pagerduty).

Recap

Interesting to focus on data model first.

Q/A

Custom database implem for state of the world. How it’s done?

Mainly python and go. Stack in go. DB itself is based on full-text query engine (comes from couchbase) reinvention of Lucene stack in go. UI is also written in go.

Fundraising, congrats. What’s next?

We are excited. Not a huge round but it’s a round. We want to make it simpler. We want “to cut the crap”. Remove the complexity for people wanting to deploy different type of applications. We see many different use cases. Trying to cover the full scope of diff apps.

Multi-cloud approach is one of our strength. Allow in the feature to let you deploy your app across many regions/cloud providers.

↑ Back to the list of talks

β˜‡ Lightning β˜‡

Nano-node - Enrico Signoriti

Future

Future of servers: 1 peta of memory per rack unit. CPU needs to catch up! Latency: reduce computational ↔ data distance. More workload in parallel New computing model: like serverless. Start a function on-demand

KV drive

Trying to solve problems but introduced even more issues

What is a nano-node?

“smaller servers”

1 node = 1 disk 3/5 watts No management Integrated CPU

Challenges

Serverless, revealed - Daniel Maher (datadog)

There is no cloud, it just someone else’s computer

Faas - Function as a service

not magic!

Event driven Architecture

Faas are build on this kind of arch

Resource utilisation

Still have to think about CPU and mem. But you also need to think about time!

Mainframes!

All clouds are basically the same

But differences:

Not a lot of difference between providers

so how to pick one?

There is no magic in serverless

still hardware constraints still software constraints even more constraints actually :)

Three heretical ideas - Mick reenman

Time series

Tech financial data moved to devops and monitoring data Now it’s everywhere! Streams from machines and applications

They are everywhere!

You’ve been told wrong things

Timeseries are Data with tags. It’s wrong.

Whatever data model you have you can narrow it to a “simple data table”. But each metric should have it’s own data model for what? Correlation, aggregations, joins

Time series has relational structure!

SQL : analysts can write queries, your devs can write apps with it....

Everyone speaks SQL tools, people… It scales really across your org

SQLish is not SQL

TimescaleDB

Timeseries database built on Postgresql Does it scale? It seems to seeing the results on the slides

20% higher inserts than Postgres

How?

Traditional transactions models is not the same as a timeseries workload. So that’s how we could optimise

Git as a continuous manager - Matthias DuguΓ©

↑ Back to the list of talks

Observability tips for HAproxy

Willy Tarreau - @WillyTarreau

Definition of observability: it indicates how well you can guess a system’s state by looking at it’s output. Wikipedia

Logs and statistics

Monitoring vs observability

Observability is important. Your client will leave you website if it doesn’t work the first time. Observability helps you with these kind of cases

LB as an observation tower

Many servers targets. So you can compare values/metrics across multiple servers. It tends to be simple and not add too much complexity.

You get a lot of logs with LBs. Which is good. Archive them.

When first incident happens: it’s usually to late and you don’t have everything you need. That’s when you go to see you LBs logs.

Failures?

You will probably see network delays on LBs. Connection retries. Connection slowdowns due to heavy firewall policies.. Client-side issues (VPN, service partner)

Metrics in HAproxy

Long terms metrics in logs.

Place unique IDs in requests to have the same trace identifier to correlate your logs

Stats page. Key specific (session/user/cookie) in your logs.

What happens during a client request

HAproxy has a lot of timers and will report duration at every event.

Recently added lua scripting so we want to report haproxy computation time (which is not the case as of today)

Timers are reported on stats page and in the logs (with lots of details). If it a request reaches a boundary you’ll get a timeout. in between you’ll get an error.

You have a set of specific termination codes which will give you information about failures (and easily scriptable due to the code failures)

HTTP statuses

HTTP status distributions will be able to show abnormal behaviours (via graphs and variations)

Queue lengths

Queues are really important to monitor. If you had only one thing to look at it’s this one!

LB fairness

If a server gets a higher amount of request = it’s misbehaving If it gets a lower amount of requests = it’s misbehaving

Error rate

Observe variations: sort them per server/per client ip/per url/per user-agent

All of these metrics are in the standard log format

You can override the log format directives.

Too much traffic to enable logs?

You can do sampling if you really want.

BUT I don’t recommend to do so. With 20k events/s it’s only 1tb/month. Not that much! If you are under that number: ENABLE LOGGING.

selective logs?

If you want to watch suspicious events.

BUT how do you determine suspicious? How do you know if your logs will give you the information you need to troubleshoot? Most of the time it’s a bad idea to do selective logging.

halog goodies

extremely useful and performant (1-2gb per second) to get response time per url/ percentiles / detect stolen CPU Get relevant information directly from your logs.

Success stories #1

Some very high percentiles measured by a client. The switch where the HAproxy was connected had TWO fibers. Due to these logs they could find that one of the fiber was not working correctly!

Success Stories #2

What happened. The server was configured to use /dev/random instead of /dev/urandom which made haproxy run out of entropy and created issues for TLS handshakes.

TL;DR

Q/A

How to make HAProxy evolve?

A company behind now. With 10+ devs. Being careful to have external contributions to make sure our ideas are not biased by our vision.

Linux kernel how was it?

Very smart people, that’s nice. When you work on old versions, you don’t need to release to often. But when you need to do, that’s a lot of work to be done!

Security list has very good people there. Exchange with maintainers for older kernels

↑ Back to the list of talks

How to demonstrate our architecture are ready?

Julio Faerman - @julioaws

So, how do you demonstrate your architecture is ready?

Not necessary a “proof” but only demonstrating to yourself (at least) “ready” is difficult to define also.

“We use a [popular stack] like..” “We are an enterprise compliant with [standard policy]..” “.. we have automated testing”

In my experience, every project mostly stops at this point and pray for the best.

Sometimes it’s enough, sometimes it’s not. Β―\_(ツ)_/Β―

Change the way of thinking about architecture: Instead of an ephemeral vision we need to change our way to look at it, Performance is not the only metric to watch for.

AWS measures architecture readiness with a “framework”:

Security automation Yay

Security is not like in the cinema: a hacker trying to enter in.

Nowadays it’s all automated.

Identity management + Encryption != Privacy

Key management is really hard.

System administrators usually have too much access…

How to demonstrate that certain data is protected (for example health data)?

Live log analysis

Serverless, what is there to protect? DDoS’ responsibility is due to the function or the provider? Dependency libraries

Predictive security operations

When data was accessed. Need to be proactive on this kind of data. After you have an incident with this kind of data you usually don’t have a second chance. It’s crucial for most businesses.

Not only for security, but also reliability

E.g. Netflix Nearly 40% (/check not sure about the number) of internet traffic. Exercise from them: Replicate traffic to be able to change regions/failures servers. Simian Army (Chas monkey).

“It’s about scaling the organisation” is similar approach for architecture design

It’s a process. You can’t just cut a monolith from one day to another.

Micro services

Infra as code you can’t live without it! It needs to be completely encoded and automated

Performance efficiency

After secure and reliable, we need fast!

128 vCPUs on a single box. HPC computers.. 4Tb of RAM on a single box.

Disk intensive? β†’ NVMe

Other components: high end GPUs, field programmable gate arrays E.g. Ryft Elasticsearch

Containers: specify just CPU and memory needs and go!

Important thing is not one or another. But the sympathy of knowing whether you need to use functions/containers/hosts.

You really need to look at your cost also! Flexibility is usually money. Performance model from your cloud providers. Lynn Langit talk (Serverless - reality or BS - notes from the trenches) NDS { Oslo }.

Operational excellence (πŸ’š)

Technique to keep things in text files under shared version control: everyone able to learn, when it was done, why it was done.

Cloud is a very democratising force! Not as useful if you don’t know what solves what and how.

We don’t want to make solutions in “boxes” (on-prem, iaas, paas, faas) but it’s better to share the experiences and knowledge about what usages are good for what provider/architecture.

More sharing of decisions. More sharing of requirements and how we make great software.

Q/A

You gave a nice list of checkbox to fill. But how to prioritise them?

That’s a good question. You can go either one by one and solve the list of requirements or get the first step of each.

Do it incrementally. Really. Do a bit of each requirements, not only because it’s how you should do but also your organisation will understand the concepts better.

Micro-services. Is it a final state or is there something even better?

Not at all. It’s a tool to grow companies. Sometimes your don’t even manage to build a good monolith, so don’t do micro-services!

It’s much more important to build your software well first.

Case after case, growing a company is hard: you need to break your company in part. And this is where micro-services can help. But it’s a tool not a goal.

↑ Back to the list of talks

Data-pipeline @Activision

Yaroslav Tkachenko - @sap1ens

Datalake / Kafka clusters for our data pipeline from consoles to our data warehouse.

1+ Pb in the datalake 600+ number of topics in our kafka cluster

10k+ messages per second. from 200b to 20kb sized messages.

This talk is not about scaling number of message or message size.

You can use best practices and guidelines available online. /”It’s the easy part”/.

But how do we scale to support to amount of games that we have? The “complexity” of topics and different messages passing through the pipeline.

The difficulty of Scalability are thus in this order in our data pipeline (from easiest to most difficult): Volumes β†’ Games β†’ Use-cases

Kafka topics partitioned designed as a distributed system highly available.

Produce/consume in parallel every partitions.

Number of topics and partitions mattered more and more with the amount of Games. Number of partitions in a fixed Kafka cluster is not infinite. Eventually you’ll reach a soft limit. Latency then goes higher in the cluster.

Scaling a kafka cluster is hard: when adding a node in an existing cluster it can take days to be ready to get all events/data.

A bit “wild west” right now for conventions. Especially topics naming.

So we use $env.$source.$title (game id).$category-$version

We allow producers to create partitions on demand. Which means every new env/title will create a new topic.

What kind of solutions to apply here?

Think about databases really. Kafka is a kind of database. A topic you can be assimilated to a table name with a database name. Let’s apply this convention on the topics we have.

With existing approach it’s “easy” to see metrics and monitor because you have all the information in the topic name. As a consumer you can consume a very specific data that you need.

BUT all this dynamic and this metadata will change (new services, deprecate titles..)

With the new approach you get a nice utilisation of topics and partitions. However it’s impossible to enforce any constraints with a topic name.

Stick to the new approach

Now we need to introduce a stream processing layer. Processing all the data and writing it back to Kafka.

Why? Well we don’t have a lot of visibility on the data. We need to understand our streams. That’s where a stream processing system is helpful to log to your monitoring system of your choice.

But it’s not an ETL (no domain/business logic here).

Filtering and routing becomes really important here. “only getting a specific title from a specific env”. A generic way to be able to have that. You should build in your processing stream layer this kind of new topics.

Refinery (internal name at Activision)

Central data team. Lots of different producers (game studios, back devs…) different format (8 different formats). When you need to consume all this kind of formats. How do you do that? Not a good idea to expose that in the topic name itself.

Trick you can use: a message envelope. Treat the message as a “byte array” and add a header with metadata to be able to route/filter your data stream. So your stream processing layer can just access these headers and don’t need to understand the business envelope. We do it with protobuf v2. Unique message ID for every message in your data pipeline. (good for deduplication for example). The “business message” is just a “byte array”. We also have the schemainfo in the metadata to understand how to read the data.

Schema registry

You need to register the schema before sending any data in the pipeline. We have a custom schema registry.

Recap

Imagine we implement all these changes. Adding new game should be as simple as adding new messages! From the data platform perspective we don’t need to do much thanks to that stream processing layer. It’s still flexible enough. It’s also easier to extend the pipeline because topics contains very precise metatype. Stream filtering helps all sort of use cases.

Q/A

Is this all live?

We still run lots of legacy games. Hard to apply all these changes consistently. But for all new games this year the message header will be in use. It’s still a work in progress :)

However the schema registry used for many years now and it helped a lot.

How important is the schema registry

For small teams: it’s not necessary. When you add more consumers/producers (diff teams / diff goals) then it becomes very important. It creates contracts between systems/teams.

Game studios can have bad reputation. How do you enjoy it?

I’m not a big video game player. But at Activision we have a very good independent model (even after acquisitions and new companies join in).

↑ Back to the list of talks

Just Right Consistency

*Marc Shapiro* - Inria

New concept. Basically you need to reconcile applications with databases.

Data as available as possible. And as consistent as it needs to be.

Your application is your top level goal to make it “correct”

Data is important

Geo distribution is important you know that.

Application logic close to your data.

The problem is when you get an update. You need to replicate those. Problem of that is the “consistency problem”. Famous CAP theorem tells you “you have to choose” between Consistent and Partition or Partition and Available.

Strong consistency

Google Spanner / Azure Cosmos DB

Round trip where you need to wait (synchronous update). Application is correct but you get a slow expensive query.

If you have a partition: you are stuck!

Eventual consistency

Opposite extreme. Do your update. Fast (as fast as your read). If there’s a partition you don’t care, you still propagate.

But you get concurrency anomalies. (updates during partitions… what do you do?)

E.g. cassandra

What’s right for your application?

Medicine FMK app: create a prescription, your doctor can add medicine in the prescription, adds the patient. Then a pharmacy can check the prescription and give you the medicine.

What do you want to be sure about?

Pattern 1

Ordering things in some order.

First you create the prescription and you fill it, THEN you add the pointers (to the doctor and the patient).

Eventual Consistency doesn’t maintain that order.

You thus need Causal Consistency it’s a way to ensure events arrive in the order that your expect. It’s AP-compatible.

Pattern 2

Joint update. All updated or none.

*”all or nothing update”*. EC doesn’t work and CP is overkill!

Pattern 3 precondition

“I have to test before giving your a medicine”. There’s an “invariant test” (the amount of medicine doesn’t go negative for instance)

Some concurrency is fine. For example if someone is ADDING a count in the medicine, then we can process and retrieve a medicine without any problem.

But if two person is retrieving at the same time the medicine: you have a problem. Precondition is not stable w.r.t concurrency. You know it in your code base directly!

TL;DR

You know the patterns in your code.

Q/A

Tell us more about AntidoteDB

Developed in academic consortium. European project. With best practices for software in mind (peer review, testings etc). It’s available on Github.

Will people move to CRDTs?

That’s the hard part: migrations for existing projects

How did you come to the CRDTs idea and how it’s used by the industry?

I was pretty much surprised the industry started using it really quick! First time in my academic career that some of my work was really useful :) We knew we could do something better for this database problem so we searched about it.

What’s next?

Focus on antidote but also focusing on the edges. Hub and spoke. We try to keep these guarantees between data layers and applications.

↑ Back to the list of talks

Kubernetes 101

Bridget Kromhout - @bridgetkromhout - slides North central of Canada Developer advocate at Microsoft

Kubernetes “From the greek: the one that steers”

Quick k8s overview. Just spoiler. How did we get to that time. And where is “here”.

Been in tech for a while. Sometimes we want to look back on old things.

History of me doing stuff with containers. Running docker in 2013 in a small company DramaFever when it was said “don’t run docker in production” before 1.0.

Containers are really great to create repeatable envs. Not really great for new failures modes. “a dev image is in prod?” WTF?

Back in the 90s we were already using containers: freebsd jails, solaris zones, LXC. They were things but not widely adopted. It used to be hard and was not that easy to use.

More usable, more accessible: Cloud Foundry, Docker, Rocket, OCI.

When you are into the containers world you believe everyone is using them. But that’s usually not the case.

When K8s came into the scene. I think it was in 2014 that I really heard about it. At that time you didn’t know it would take all the hype.

It is a step. But it’s *not the final state,* really. There’s lots of other orchestrators in production that work fine too.

Remember it’s a TOOL not a GOAL.

What people are excited about in that “landscape”. What they have in common: focus on containerisation (isolation, reproducability, dynamic orchestration). Focus on micro-services. Micro services are great but not a magical solution for every thing.

Ecosystem scene is complex: because the problem itself is complex!

/”Computers are hard and when you add people it’s even harder and then it goes on fire”/

Your just gonna move the complexity somewhere else.. You get to chose where the problems can be.

Kubernetes help you to manage that complexity.

From scratch tutorials

Architecture

API, Control plane (master nodes), worker nodes

Built-in service discovery.

where are we gonna spend our energy and efforts?

Master architecture

API server, Scheduler, controller based on etcd

The pretty opinionated thing in K8s: ETCD. Everything else is pretty much not opinionated.

Node architecture

Pod can have one or more containers in it. Shared resources. Pods can be scaled. Optional addons: DNS, UI, …

How to learn about K8s?

Kubernetes workshop doesn’t give you all the tools to go back in your company and build everything there!

How to democratise the learning about K8s? Open source K8s training of course! JΓ©rome Petazzoni tries to do so: jpetazzoni/container.training

Walk through concepts and terms, walk through operation k8s.

Getting people to use ~kubectl,~ getting people to use containers, what to do with a registry.

K8s dashboard, how to get it securely exposed.

There’s direction for Azure and AWS for now.

Always emotion is future. We have upcoming events coming. Run the training at events! Send us PRs

Lucas Kaldstrom’s intro to kube

Projects to watch: Helm, Brigade, Virtual Kubelet (plug your hosts in a K8s cluster, or anything else (VM, on-prem) to be treated as a kubernetes node)

Erik St Martin, “k8s is not the thing, it’s what’s going to get us to the thing”

I work at microsoft and my job is to get people use linux. Join us we are hiring.

Q/A

Devopsdays tells us about that

2015 I took over looking after the community. Lots of conferences over several continents. It’s a great thing. Do participate if you can and if you want to organise when come tell me!

More about Microsoft? The involvement in open source is striking

It’s clear, where the growth is. It’s exciting AND relevant to business. And lots of customers are coming saying they want to use Linux… Putting investments into what clients want to use is an easy choice to make!

It’s exciting time! People sharing their experience with each others. Trying to serve the public and learn from.

I unfortunately couldn’t stay for the last two talks, SΓ©bastien Elet did a great job at summarising the talks so I leave you to his tweets

↑ Back to the list of talks

At Scale everything is Hard

Paul Dix - @pauldix

https://twitter.com/SebastienElet/status/1002576816442363905

↑ Back to the list of talks

Infrastructure Data

Jeremy Edberg - @jedberg

https://twitter.com/SebastienElet/status/1002582998116651008

↑ Back to the list of talks