Added shuffle sharding support to generate a subring #3090

pracucci · 2020-08-26T11:21:54Z

What this PR does:
The current Subring() implementation was conceived as an experimental shuffle sharding support in Cortex. It works fine in regards to build a subring of N nodes and guaranteeing stability and consistency but unfortunately suffers two issues:

It doesn't shuffle instances between tenants
It's not zone aware (the resulting subring may be very inbalanced in regards to zones)

In this PR I proposed an alternative ShuffleShard() implementation which is inspired by AWS Route53 infima library and guarantee stability and consistency, while offering nodes shuffling and zone awareness.

Will be done in subsequent PRs:

Introduce a cache for each tenant subring
Documentation

Which issue(s) this PR fixes:
N/A

Checklist

Tests updated
Documentation added
CHANGELOG.md updated - the order of entries should be [CHANGE], [FEATURE], [ENHANCEMENT], [BUGFIX]

tomwilkie · 2020-08-26T12:58:17Z

/cc @thorfour (original author of #1947) and @harry671003, who's been poking around in this area.

pkg/ring/ring.go

tomwilkie · 2020-08-26T13:02:27Z

pkg/ring/ring.go

Just want to check: in the case where we're not using zones, zone = ingesterID. Will this algorithm end distributing all users to all ingesters in this case?

in the case where we're not using zones, zone = ingesterID

It was in the initial implementation but was soon fixed in #2404. When not using zones, the zone is an empty string, which works fine here (all ingesters belong to the same zone).

However, I'm wondering if we should add a boolean flag to the ring to explicitly enable zone awareness, disabled by default instead of just relying on the fact zones are set (think about a rolling update adding zones to ingesters, you may want to rollout all ingesters first and then enable it).

My 2 cents is if it doesn't add too much complexity, it's probably better to favor rollout safety for changing the flag.

Agree @ranton256. Will do it in a separate PR (since it's unrelated from the work done in this PR).

thorfour · 2020-08-26T13:58:49Z

It doesn't shuffle instances between tenants

I guess I don't understand this comment. Can you elaborate on what this problem is?

pstibrany

Great work!

pkg/ring/ring.go

pstibrany · 2020-08-26T14:30:46Z

pkg/ring/ring.go

What do you mean by "probabilistic hashing"?

Uniqueness is not guaranteed, but it offers a low collision probability. What we're using here is a probabilistic data structure.

pracucci · 2020-08-26T14:54:29Z

It doesn't shuffle instances between tenants

I guess I don't understand this comment. Can you elaborate on what this problem is?

If we pick nodes following the "next" tokens in the ring (like Subring() is working right now) we pick consecutive nodes instead of shuffling them. Think about 2 tenants: one whose initial hash match the node X in the ring, and the second matches the node X+1 in the ring. They will share all nodes minus 1 in the subring. Shuffle sharding is about shuffling nodes to make this case probabilistically unlikely to happen, which is what this PR does.

thorfour · 2020-08-26T15:16:03Z

It doesn't shuffle instances between tenants

I guess I don't understand this comment. Can you elaborate on what this problem is?

If we pick nodes following the "next" tokens in the ring (like Subring() is working right now) we pick consecutive nodes instead of shuffling them. Think about 2 tenants: one whose initial hash match the node X in the ring, and the second matches the node X+1 in the ring. They will share all nodes minus 1 in the subring. Shuffle sharding is about shuffling nodes to make this case probabilistically unlikely to happen, which is what this PR does.

So is the value of this for really low tenant counts to make it more likely to evenly distribute load? Because the example of 2 tenants overlapping will always be true at large enough tenant counts.

pracucci · 2020-08-26T15:26:56Z

So is the value of this for really low tenant counts to make it more likely to evenly distribute load?

No. It works very well with a large number of tenants too (assuming a reasonably large number of nodes too, but order of magnitude smaller then tenants).

Because the example of 2 tenants overlapping will always be true at large enough tenant counts.

It depends on the cluster size. If you have a cluster with 100 ingesters and you randomly pick 6 nodes for each tenant, you end up with 75M possible combinations, and the chances to have overlapping nodes (with shuffle sharding) are:

Probability of 6 equal instances: 0.00000008%
Probability of 5 equal instances: 0.00004731%
Probability of 4 equal instances: 0.00550018%
Probability of 3 equal instances: 0.22489615%
Probability of 2 equal instances: 3.83729063%
Probability of 1 equal instances: 27.62849251%
Probability of 0 equal instances: 68.30377314%

I created this spreadsheet to easily compute it.

For more information, you could take a look at this and the reference implementation linked in the PR description.

thorfour

@pracucci thanks for the explanation! LGTM

ranton256 · 2020-09-01T15:32:13Z

docs/configuration/config-file-reference.md

+# in the per-tenant overrides, a value of 0 disables shuffle sharding for the
+# tenant.
+# CLI flag: -distribution.ingestion-tenant-shard-size
+[ingestion_tenant_shard_size: <int> | default = 0]


I take it based on the similar "store_gateway_tenant_shard_size" flag from https://github.com/cortexproject/cortex/pull/3069/files you are planning to name all the similar flags "_tenant_shard_size"?

Also, are you planning to take/keep the -experimental off this when you are ready to submit?

I take it based on the similar "store_gateway_tenant_shard_size" flag from https://github.com/cortexproject/cortex/pull/3069/files you are planning to name all the similar flags "_tenant_shard_size"?

Yes, I do. The reason is that overrides are specified within the same YAML node, so we need a way to clearly differentiate them. What do you think?

Also, are you planning to take/keep the -experimental off this when you are ready to submit?

I would plan to remove the experimental flag once ready to submit, unless you have any concerns. Generally speaking, I think originally adding the experimental prefix to CLI flags was a mistake and shouldn't be done anymore. What do you think?

I think you mean document as experimental but not name the flag experimental? That sounds fine to me.

Yes Richard, you're correct (sorry for being unclear). This way, once we're all comfortable marking it stable, it will just be a doc change instead of a config change.

ranton256

LGTM

pstibrany · 2020-09-08T11:10:51Z

There is additional work (caching of user rings) available in https://github.com/pstibrany/cortex/tree/fix-shuffle-sharding branch, not yet tested.

Signed-off-by: Marco Pracucci <[email protected]>

pstibrany

Great job!

Signed-off-by: Marco Pracucci <[email protected]>

…ortex#3090) * Added shuffle sharding support to generate a subring Signed-off-by: Marco Pracucci <[email protected]> * Solved all TODOs Signed-off-by: Marco Pracucci <[email protected]> * Simplified unlock Signed-off-by: Marco Pracucci <[email protected]> * Fixed linter Signed-off-by: Marco Pracucci <[email protected]> * Added benchmark Signed-off-by: Marco Pracucci <[email protected]> * Replaced Subring() with ShuffleShard() Signed-off-by: Marco Pracucci <[email protected]> * Small improvements Signed-off-by: Marco Pracucci <[email protected]> * Shortened CHANGELOG entry Signed-off-by: Marco Pracucci <[email protected]> * Make golang doc happy Signed-off-by: Marco Pracucci <[email protected]> * Fixed flag name and added integration test Signed-off-by: Marco Pracucci <[email protected]>

pracucci requested review from tomwilkie and pstibrany August 26, 2020 11:21

pull-request-size bot added the size/XL label Aug 26, 2020

tomwilkie reviewed Aug 26, 2020

View reviewed changes

pkg/ring/ring.go Outdated Show resolved Hide resolved

tomwilkie reviewed Aug 26, 2020

View reviewed changes

pstibrany approved these changes Aug 26, 2020

View reviewed changes

thorfour approved these changes Aug 26, 2020

View reviewed changes

pracucci force-pushed the fix-shuffle-sharding branch from 0629048 to ea4c258 Compare August 27, 2020 07:52

ranton256 reviewed Sep 1, 2020

View reviewed changes

pracucci added 7 commits September 17, 2020 14:56

Added shuffle sharding support to generate a subring

00b0600

Signed-off-by: Marco Pracucci <[email protected]>

Solved all TODOs

485ceb6

Signed-off-by: Marco Pracucci <[email protected]>

Simplified unlock

115c735

Signed-off-by: Marco Pracucci <[email protected]>

Fixed linter

162569f

Signed-off-by: Marco Pracucci <[email protected]>

Added benchmark

baeeed2

Signed-off-by: Marco Pracucci <[email protected]>

Replaced Subring() with ShuffleShard()

6b09381

Signed-off-by: Marco Pracucci <[email protected]>

Small improvements

7131ffc

Signed-off-by: Marco Pracucci <[email protected]>

pracucci force-pushed the fix-shuffle-sharding branch from 3783b39 to 7131ffc Compare September 17, 2020 13:08

pracucci marked this pull request as ready for review September 17, 2020 13:08

pracucci added 2 commits September 17, 2020 15:15

Shortened CHANGELOG entry

9df4eac

Signed-off-by: Marco Pracucci <[email protected]>

Make golang doc happy

fe24ed2

Signed-off-by: Marco Pracucci <[email protected]>

pstibrany approved these changes Sep 17, 2020

View reviewed changes

Fixed flag name and added integration test

048341a

Signed-off-by: Marco Pracucci <[email protected]>

pracucci merged commit c0d3995 into cortexproject:master Sep 17, 2020

pracucci mentioned this pull request Sep 18, 2020

Added flag to explicitly enable zone-awareness replication and added store-gateway support #3200

Merged

3 tasks

pstibrany mentioned this pull request Sep 21, 2020

Caching of shuffle-sharded subrings #3214

Merged

3 tasks

Added shuffle sharding support to generate a subring #3090

Added shuffle sharding support to generate a subring #3090

Uh oh!

Conversation

pracucci commented Aug 26, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

tomwilkie commented Aug 26, 2020

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

thorfour commented Aug 26, 2020

Uh oh!

pstibrany left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pracucci Aug 27, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pracucci commented Aug 26, 2020

Uh oh!

thorfour commented Aug 26, 2020

Uh oh!

pracucci commented Aug 26, 2020

Uh oh!

thorfour left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ranton256 left a comment

Choose a reason for hiding this comment

Uh oh!

pstibrany commented Sep 8, 2020

Uh oh!

pstibrany left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

pracucci commented Aug 26, 2020 •

edited

Loading

pracucci Aug 27, 2020 •

edited

Loading