Containers-Microservices-Node.txt

NodeSummit, San Francisco, CA, Wed 27 Jul 2016

Containers, Microservices, Node.js

Jacob Groundwater, Engineering Manager, New Relic
@0x604

Justin Beckwith, Product Manager, Google
@JustinBeckwith

Tim Gross, Product Manager, Joyent
@0x74696d

Naresh Sikha, Enterprise Architect, StubHub

Are things getting too complicated?
Don't try to eat the whole thing on day one.
Get more sophisticated over time.
Microservices is a technical solution to a cultural problem.
Google has been using containers for years, they just weren't Docker containers.
Tooling.
Layered infrastructure, incremental.
Docker solves the 'it works on my machine' problem.
Containerizing a codebase you understand first is a good way.
Be honest about issues that will occur.
Cognitive load and deciding how far to go.
Container has to live in broader context.
Uber switching from PostgreSQL to MySQL.
Motivation, devs don't understand failover at app level.
Need uptime plan for routing (Google).
Need common exception handling, common logging (New Relic).
At Google, every request is tracked by unique ID assigned by the load balancer.
Propagate ID throughout infrastructure.
Using gRPC to communicate at Google.
Google already has rich lib support for this.
At Joyent:
  - Infrastructure team provides base layer.
  - App team adds app layer.
  - Monitoring team adds monitoring layer.
A monolithic deploy of microservices hasn't really changed anything.
How do we test the stack as a whole?
No generic tools for doing this.
Netflix uses fault injection testing.
Curated test environments, white gloves.
Lots of people take database out of containers.
Lots of container infrastructures have terrible I/O.
Model a lot of people have come up with, over the network, block storage.
But new set of failure modes.
Architecture - does app need to be aware of replication?
Replication at block layer opens up Byzantine failure modes.
But it gets it working at app level.
Avoids costs of expensive replication hardware.
Assume things will go down.
Five nines times several times degrades overall SLA.
Problem isn't services with low reliability, because you expect it, code around it.
Race failures are the problem.