[Cache] Improve RedisTagAwareAdapter invalidation logic & requirements

[andrerom]

Q	A
Branch?	4.4
Bug fix?	yes, and improvment
New feature?	no
BC breaks?	no
Deprecations?	no
Tests pass?	yes
License	MIT
Doc PR

Changes logic of invalidation in RedisTagAwareAdapter in order to:

• Delete the tag key on invalidation => avoiding possible left behind empty tag keys that Redis is not allowed to evict, gradually consuming slightly more memory

Positive side effects of no longer using sPOP:

• Lowered requirements to Redis 2.8, and no specific version constraint for phpredis
• Lift limitation of 2 billion keys per tag (Now only limited by Redis Set datatype: 4 billion)

[nicolas-grekas]

That's cool!
A hypothetical future step could be to schedule the cleanups using Messenger :)

[andrerom]

Some setbacks here, clients are not so complete:

• RedisArray does not support SSCAN, you'll get:
  Parameter 2 to Redis::sscan() expected to be a reference, value given
• Predis alos have limited SSCAN and gives the following if key does not exists:
  Predis\Response\ServerException: ERR no such key

These can potentially be fixed by instead using SMEMBERS + skipping SSCAN/SMEMBERS on tagId if RENAME returned 0 (not found in this case).

More of a Server issue:

• Cluster on Predis and RedisCluster does not support RENAME:
  RedisCluster::rename(): Keys don't hash to the same slot

This one I'm unsure of, should we try to make sure we RENAME within same node?

Suggestions @nicolas-grekas / @berezuev?

[nicolas-grekas]

Can't we rename to a key that will be guaranteed on the same node, using the special syntax they support for this?

[andrerom]

using the special syntax they support for this?

Do you happen to remember where that is described? Thought about that too, but didn't know where to look.

[nicolas-grekas]

Sure, see https://redis.io/topics/cluster-spec, the section about "hash tags"

[stof]

is it possible to have { and } in our normal tagId Redis keys ? If yes, we won't be able to use hash tags to achieve our goal.
Otherwise, this can indeed by used, making the temporary tag id being '{'.$tagId.'}'.$randomSuffix instead of $tagId.':'.$randomSuffix.

[nicolas-grekas]

The alternative could be to talk directly to the corresponding node in the cluster. That's possible AFAIK.

[nicolas-grekas]

I think there is a function for that :)
Happy 🌴 !

[andrerom]

RedisCluster failure was fixed in be04079. I did not find a simple way to deal with RENAME on Predis across RedisCluster & PredisCluster so left that as-is, could instead use SPOP there if prefered.

AppVeyor failure seems unrelated (HttpClient\Tests\NativeHttpClientTest::testNotATimeout), so rebased and moving this to review.

[nicolas-grekas]

please rebase, you've been caught by fabbot :)

[andrerom]

let's throw for now in this situation and open an issue on Predis?

last commit 2017...

[nicolas-grekas]

Another reason to throw and encourage using phpredis...
Let's still create the issue for the record...

[fabpot]

Mergeable?

[andrerom]

IMHO Yes, we can do the improvements as a follow up, it will anyway need some investigation and testing. But it's entirely up to you two.

[nicolas-grekas]

I'd prefer doing it in this PR, depending on your availability to do it :)

[nicolas-grekas]

Hello, I just push-forced this here:
https://github.com/symfony/symfony/compare/db80b9db0519ec73165b9bc2d57674993c7a3d5e..28948eeb8b

Should fix the last comment.

[andrerom]

+1 Looks good to me 👍

Further improvements here depends on future improvements in Predis if anyone is up for the tasks:

1. Exception on rename on missing key (should align with phpredis: false response when missing)
2. Lack of api to find slot of key when using RedisCluster class (should also be added to ClusterInterface, based on implementation in PredisCluster class)

[nicolas-grekas]

Exception on rename on missing key

Actually, I just discovered this is already possible! PR updated to use pipelining.

Lack of api to find slot of key when using RedisCluster

that's a bit more involving, as that'd require implementing the logic to react to MOVE responses. Should be handled at the Predis level instead, as phpredis does.

[nicolas-grekas]

Thank you @andrerom.