Scaling Datastores at Slack with Vitess

https://slack.engineering/scaling-datastores-at-slack-with-vitess/

by talonx • 6 years ago

99 34 6 years ago

wiredfool2• 6 years ago

Sharding based on customer is an attractive way to get to shards, potentially with minimal interference in ops or data model, but truly sucks when your biggest customers don’t fit on one shard anymore.

kamikaz1k• 6 years ago

Has that ever happened to you? Would love to read about that story and hopefully the solution...

juancampa• 6 years ago

The article addresses this particular point

Each keyspace is a logical collection of data that roughly scales by the same factor — number of users, teams, and channels. Say goodbye to only sharding by team, and to team hot-spots!

sougou• 6 years ago

The way to reason about this is: If a customer had a huge business, what key would they shard by? Then choose that as the sharding key for all customers.

philsnow• 6 years ago

sougou is one of the original vitess devs, he would know this very well.

Hi Sugu!

atombender2• 6 years ago

The solution is to have different shard sizes. Your largest customer needs to fit on the largest shard size (extra large VMs or whatever).

wiredfool2• 6 years ago

Fins, but there comes a time when your largest customer doesn’t fit on your largest machine.

I wouldn’t say that killed my first startup, but it certainly didn’t help.

chevman• 6 years ago

Yeah, seems like the key point here is be careful architecting your infrastructure around a somewhat arbitrary construct like 'customer'. The relevant entity to scale around for Slack is likely 'active users' or even 'messages'. Which may map loosely to 'customers', but that's almost incidental.

vp89892• 6 years ago

It looks like Azure has now integrated Citus into it's Postgres offering, but Im curious why AWS hasn't integrated either Vitess or Citus at all?

It seems like a well executed combination of Vitess/Citus + read replicas pointing to decoupled page storage would result in a fully featured non-proprietary database that scales all the way down to your laptop and all the way up to the heaviest OLTP use cases. If that existed, only truly masochistic people would continue using things like DynamoDB.

marcinzm• 6 years ago

If that existed, only truly masochistic people would continue using things like DynamoDB.

Which is probably why AWS doesn't support it, they make less money and have less lock in on their clients.

Scarbutt2• 6 years ago

They want you to use Aurora instead.

carlineng• 6 years ago

Citus was acquired by MSFT in 2019, hence the integration. Vitess is primarily a MySQL-based solution, and I imagine that AWS would rather have its customers on Postgres-flavored DBs, given that MySQL is currently owned by Oracle.

perfectspiral• 6 years ago

The Citus extension is still open-source though, right? Although I think it was a solid strategic move by MSFT to acquihire the Citus devs (probably all these people were given substantial golden handcuffs) to block AWS from hiring experts to integrate this extension into their PG offering.

mosselman• 6 years ago

Is there something comparable for PostgreSQL that makes it easy to setup clusters with? Maybe with automatic fail overs and HA?

Scarbutt2• 6 years ago

Not a library but you can achieve the same with less work using amazon aurora postgres.

vp89892• 6 years ago

It's not the same, you can't scale out from 1 writer on Aurora.

isatty• 6 years ago

And a boat load of money

shlomi-noach• 6 years ago

Deepthi, tech lead for Vitess, just published this RFC a few days ago: https://github.com/vitessio/vitess/issues/7084

Vitess plans to support PostgreSQL, but there is no work on that at this time.

manquer• 6 years ago

Citus does sharding as an extension for postgres.

If all you want is HA and FO there are simpler solutions like pgpool and pgbouncer etc.

dilyevsky• 6 years ago

CockroachDB uses pg protocol but doesn’t support all the features (close enough though).

nkcmr• 6 years ago

Not a DBA, but this seems comparable:

stolon - PostgreSQL cloud native High Availability https://github.com/sorintlab/stolon

merb• 6 years ago

on k8s there is:

- crunchydata - the thing crunchy is based on (well actually they are more based on the underlying things like patroni): zalando postgres operator https://github.com/zalando/postgres-operator, really really solid stuff

yashap2• 6 years ago

Today, we serve 2.3 million QPS at peak. 2M of those queries are reads and 300K are writes.

Damn, even for a service as popular as Slack, that’s significantly more than I expected. Slack has ~12-13 mil DAUs, right? I assume at peak time of day, maybe 2-3 million actively using the product at the same time? If that’s a fair assumption, that’s roughly 1 MySQL query per second, per active customer - seems like a fair bit? I wonder if they do polling (instead of websockets)?

At my work we have roughly 1 order of magnitude fewer DAUs, but roughly 2 orders of magnitude fewer QPS. And we also have chat areas of the product.

karlding• 6 years ago

Slack has ~12-13 mil DAUs, right?

As of September 2019, Slack reported 12 million DAUs [0]. Of course, that is pre-pandemic, and the only figures that Slack has provided regarding demand in 2020 are these tweets [1] from Slack CEO Stewart Butterfield back in March. In those tweets, it mentions that Slack was serving 12.5M simultaneously connected users.

[0] https://slack.com/blog/news/work-is-fueled-by-true-engagemen...

[1] https://twitter.com/stewart/status/1243000487365861376

jhgg• 6 years ago

It is a pretty high QPS. I work on a chat service that is quite popular (even more popular than Slack) and looked at our QPS to our persistent data stores. It is less than an order of magnitude compared to slack's when you normalize for DAU. Really wonder what they're doing over there.

derekperkins2• 6 years ago

Whatever the service, it isn't the platform that Slack is. I would guess that a huge percentage of their db traffic is from apps and custom integrations.

ahachete2• 6 years ago

GitLab.com is serving around 300QPS at peak (see https://about.gitlab.com/blog/2020/09/11/gitlab-pg-upgrade/).

While that's like 1/7th of Slack's traffic, it is served by a single PostgreSQL cluster.

That's the comment.

yashap2• 6 years ago

May want to edit your post - I believe you meant 300K QPS, not 300 QPS :) But yeah, that’s very impressive to handle that much load without sharding! Though it is a cluster of 12 instances, each with 96 CPUs and 614 GB RAM - those are some seriously beefy machines.

Slack are getting 300K WRITES per second, though. Which does mean you have really no choice but to shard - read replicas are nice for scaling reads, but do nothing for writes, and that’s A LOT of writes.

conradfr• 6 years ago

Does that account for all those app/bots monitoring/CI/VCS/etc that "spam" channels?

Scaling Datastores at Slack with Vitess

Related Stories