Does ZRANGE BYLEX behaviour differ from Redis?

[yatriks]

Perhaps this is covered somewhere else or documented, but I see a variance in the behaviour of ZRANGE BYLEX between redis and kvrocks. Example:

❯ redis-cli -p 6666 zadd test 1 abcd0 2 abcd1 1 abcc0 2 abcc1
(integer) 4

❯ redis-cli -p 6666 zrange test \[abcd0 \[abcd5 bylex
1) "abcd0"
2) "abcd1"

❯ redis-cli -p 6379 zadd test 1 abcd0 2 abcd1 1 abcc0 2 abcc1
(integer) 4

❯ redis-cli -p 6379 zrange test \[abcd0 \[abcd5 bylex
1) "abcd0"
2) "abcc1"
3) "abcd1"

My compose file is like so:

services:
  kvrocks:
    image: 'apache/kvrocks:latest'
    ports:
      - '6666:6666'
  redis:
    image: 'redis:8-alpine3.23'
    ports:
      - '6379:6379'

I believe this has to do with the behaviour of the internal rocksdb column family keys?

I find the kvrocks behaviour to be more intuitive but I don't know enough about lexicographical ordering to determine if it is correct or wrong. There are other range queries that have matching answers. There are times where the answer I expect is the one kvrocks provides, as it is able to use BYSCORE for score queries and BYLEX for lex queries which ignore score which is useful, but again, unsure if that is correct.

Please let me know if there's any other information I can provide, or if there's anything I've missed.

[git-hulk]

@yatriks It's a really good finding.

When all the elements in a sorted set are inserted with the same score, in order to force lexicographical ordering, this command returns all the elements in the sorted set at key with a value between min and max.

I guess that's a design constraint inside Redis, because it now doesn't have an extra member name index to walk through when ranging with lex.

See [1].

I find the kvrocks behaviour to be more intuitive but I don't know enough about lexicographical ordering to determine if it is correct or wrong. There are other range queries that have matching answers.

Yes, I also agree that Redis behavior was a bit counterintuitive, but I don't have a clear answer for the correctness. Keeping the same behavior with Redis would have less confusion from the user side, but it would be counterintuitive.

To see @apache/kvrocks-committers have any thoughts on this?

[1] https://redis.io/docs/latest/commands/zrangebylex/

[jihuayu]

This has been the case for a long time without issue, maybe the current implementation is what users expect. We could stick with it and just note this difference in the docs, reopening the discussion only if someone objects later.

[git-hulk]

Yes, I also strongly recommend keeping the current behavior because I believe it's a technical design trade-off in Redis rather than a "design intent".

[jihuayu]

Hi @yatriks. I found that our docs don't mention this. Could you help us update them? https://kvrocks.apache.org/docs/supported-commands#zset-commands

[yatriks]

What do you think of appending, to each of ZRANGE, ZLEXCOUNT, ZRANGEBYLEX, ZRANGESTORE, ZREVRANGEBYLEX, ZREMRANGEBYLEX:

Lexicographical order in kvrocks disregards scores; this may differ from Redis behavior when members have varying scores.

[jihuayu]

No problem, you can add a hyperlink to this text that jumps here.

[yatriks]

submitted: apache/kvrocks-website#373

[jihuayu]

Thanks