A talk on how to think about choosing a distributed messaging technology, and some notes on how to avoid locking yourself into a single choice, keeping your platform able to grow as needs change.
2. A little about myself
● Sean Kelly
○ Also known as Stabby
● I went from .NET to Ruby to Go
○ But my favorite language is SQL
● Core maintainer of Tapjoys:
○ Chore - https://github.com/Tapjoy/chore
○ Dynamiq - https://github.com/Tapjoy/dynamiq
● I love IPAs
3. Speaking of Tapjoy...
We do…
● 1.8 Billion Requests per minute
○ And almost as many messages per day
● ~170 Million Jobs per day
● All on ~750 EC2 instances and private
servers
● A stocked double-kegerator
○ Right now: Pinner IPA / Cisco Summer of Lager
5. Messaging is...
● A way to share important events, without
needing to know who's listening
● A way to handle processing events and
information at a larger scale
● Not all that unlike “background jobs”
○ Jobs: “I’ll do this later”
○ Messaging: “Other people will do this later”
9. Messaging and You
Now you have several services, and they all
need to share info
Monolith
1.5
Jobs
One-off which
becomes a
core part of
your business
Jobs
Failed
attempt at
Micro
Service
Reporting
System
10. Sure, but how can you
actually use Messaging?
Those weren’t even very good drawings
They didn’t have lines or anything
11. What types of Messaging are there?
● 1:1, traditional “Queueing”
○ Basic push / pull model of doing work
○ Common with asynchronous job processing
○ RabbitMQ, ActiveMQ, SQS, Disqueue, Dynamiq, NSQ
● Fanout
○ Broadcast style publishing, all listeners get a copy
○ Ex: A game pushing out notifications of an update
○ Most technologies with 1:1 queues support this in some way
12. What types of Messaging are there?
● Routing
○ Intelligent fanout, routes to listeners based on message
metadata
○ Newsgroups: Subscribe to food.charcuterie.*, get bressola
○ RabbitMQ does this pretty well
● Streaming
○ Persistent connection, constant source of raw bytes
○ Twitter's Firehose is one example
○ Kafka is a current popular choice
○ Really popular with the Scala / Spark crowd
13. OK, so my Apps and
Services need to talk
Can’t I just stick it all in a shared database
and be done with it?
15. Why not just stick it all in a DB?
● You can some share of your data this way
○ Depends on the use case, type of information
○ This is outside the scope of this talk
● Databases are not designed for delivering
messages
○ Any “queue” tables will be ridiculously contended
○ No atomic “pull” options
16. So, what does Tapjoy do?
You guys must have solved this, right?
17. At Tapjoy, we use...
● RabbitMQ
○ Moves analytics events to reporting endpoints by way of complex filesystem / s3 approach
○ Single node with sharded queues
○ Rabbit HA cannot handle our scale
● SNS / SQS
○ SNS in some newer projects, mostly for fanout
○ SQS for all traditional background jobs
● Kinesis
○ Pilot integration for a new analytics pipeline
○ Being supplanted with Kafka
● Kafka
○ New analytics pipeline
○ Used to distribute metrics to both the new endpoint as well as the existing one for
verification
● Dynamiq
○ Inhouse Open Source SNS/SQS-alike built on top of Riak 2.0
○ Currently used to circumvent complicated and slow legacy messaging service
18. But I’m not really here to
talk about Tapjoy
Not entirely
I’m more interested in you
19. So, what do I pick?
There are so many choices, and they all
seem like they’d work
20. I’m not really here to tell
you what to pick, either!
I’d rather talk to you about how to pick, and
how you integrate your choices
Distributed Systems are all about tradeoffs
21. Ask: What are my actual needs?
● Planning for 2 years down the road is smart
○ But solutions right now get shit done
○ Include a cost projection with scale estimates
● Build a prototype (or two)
○ Try to iterate quickly
○ Understand how you’d use whatever you choose
○ Don’t be afraid to move on
○ Look at multiple client libraries
■ Look for: Docs, Active repos, Idiomatic
22. Ask: What is my latency tolerance?
● Publishing Messages
○ How much time can your app tolerate for publishing?
○ What does publish latency look like during an issue?
○ Consider the worst-case scenario when planning
● Consuming Messages
○ Can you run multiple consumers without impacting
the service?
● End to End
○ How fast is the whole experience, round trip?
23. Ask: What level of durability?
● Client
○ Batched VS Unbatched / Streaming
○ Acknowledged writes
● Server
○ Messages held in memory VS disk
○ Messages highly-available?
○ Recover from network partitions safely?
○ At-Most-Once VS At-Least-Once
■ Exactly-Once is something of a myth
24. Ask: What about throughput?
● How many producing clients do you have
● How many messages per second will they submit
○ Does message size impact performance?
● What size should the cluster be?
○ Super cluster VS specialized clusters
● How many consumers it takes to keep pace
○ With room to grow
25. Ask: What does failure mean?
● What does a message publishing error
mean?
● What does a delay in the processing pipeline
mean?
● What does a “lost” or failed message mean?
● What does a total failure of the messaging
system mean?
26. Ask: What behavior do I want?
Is it…
● CA?
○ Not distributed, will be difficult to scale past 1 box
○ Traditional RDBMS systems are typically CA
● CP?
○ Good if you need strongly consistent data
○ Partitions can cause data unavailability
● AP?
○ Good if you need complete availability
○ Eventual consistency can often be “good enough”
27. Okay, so I lied a little bit
I’ll give you one recommendation
28. Do you have...
● Relatively small (< 256kb) message sizes?
● Not so strict (~50ms) latency requirements?
● Throughput on the order of 100m or less per
month?
● A tolerance or capability to handle the
occasional duplicate message?
● No concern around being locked into a
vendor-specific technology?
29. Go use SNS and SQS
immediately
Leave here now and just do it
It’s easy, it’s cheap (at that scale), and you
don’t need to maintain it
30. Ok, so I picked
“something”
Anything else to know?
31. You don’t have to choose just 1
● It’s a falsehood that you need 1 perfect
technology
○ Each has strengths, weaknesses, and ideal use
cases
● Don’t be afraid to use something else
○ If you’re lucky, your app lives long enough to see
many different infrastructure needs
32. Avoid direct implementations
● Wrap the notion of Publishing in an interface
○ Most technologies look surprisingly similar to publish
○ You can wrap this in a simple interface, and switch
implementations as needed
● Consuming is usually unique per technology
○ Just write a new one
○ Trying to interface this part is probably more trouble
than it’s worth
○ Play to the unique strengths of the technology
33. Interfacing your Messaging choices
● Sending messages is often as simple as a name and a chunk of
data
○ Define a simple interface for pushing arbitrary data towards a
named endpoint
○ A name and a string of JSON is usually enough to get going
○ At Tapjoy, we use our Chore library to handle abstracting
message publishing from messaging technologies
● Destinations are independent from messages
○ You could need to switch sending messages to a new
technology
○ You could have 2 or more different systems depending on the
information in a given message
34. How do I change messages safely?
● Wrap messages in a simple envelope
○ Keep metadata about the message distinct from metadata
about the event it describes
● Define schemas for message bodies
○ Schemas give you a catalogue of message definitions, and the
ability to version them
○ At Tapjoy, we use our TOLL to build endpoint-agnostic clients
based on schemas, and register them to use Chore publishers.
● Consumers need older schemas
○ Lets them reason about how to handle older messages
○ Keep a backlog of N older versions, drop support for > N
36. Keep in mind
● Distributed Systems - all about tradeoffs
○ Never trade “P”
● Understand your needs
○ Latency, Throughput, Availability, Durability
● Understand how it fits into your architecture
● Interfaces are your friend
○ They can give you a lot of flexibility
37. Keep in mind
● Use schemas and versioning to support changes to
messages themselves
● Just pick something
○ Build a prototype, or two (or three)
○ Your second try will probably go better
○ SNS/SQS is a decent choice, if latency isn’t a
concern
● Tapjoy is a great place to work on these kinds of
problems at huge scale