Hashing used in Citus

[Rohit25negi]

What is the Hashin technique used in Citus. Normal modulo hashing or consistent hashing?

[marcocitus]

Consistent hashing.

If you look at pg_dist_shard you see the range of hash values for each shard:

select * from pg_dist_shard where logicalrelid = 'test'::regclass;
┌──────────────┬─────────┬──────────────┬───────────────┬───────────────┐
│ logicalrelid │ shardid │ shardstorage │ shardminvalue │ shardmaxvalue │
├──────────────┼─────────┼──────────────┼───────────────┼───────────────┤
│ test         │  102008 │ t            │ -2147483648   │ -2013265921   │
│ test         │  102009 │ t            │ -2013265920   │ -1879048193   │
│ test         │  102010 │ t            │ -1879048192   │ -1744830465   │
│ test         │  102011 │ t            │ -1744830464   │ -1610612737   │
│ test         │  102012 │ t            │ -1610612736   │ -1476395009   │
│ test         │  102013 │ t            │ -1476395008   │ -1342177281   │
│ test         │  102014 │ t            │ -1342177280   │ -1207959553   │
│ test         │  102015 │ t            │ -1207959552   │ -1073741825   │
│ test         │  102016 │ t            │ -1073741824   │ -939524097    │
│ test         │  102017 │ t            │ -939524096    │ -805306369    │
│ test         │  102018 │ t            │ -805306368    │ -671088641    │
│ test         │  102019 │ t            │ -671088640    │ -536870913    │
│ test         │  102020 │ t            │ -536870912    │ -402653185    │
│ test         │  102021 │ t            │ -402653184    │ -268435457    │
│ test         │  102022 │ t            │ -268435456    │ -134217729    │
│ test         │  102023 │ t            │ -134217728    │ -1            │
│ test         │  102024 │ t            │ 0             │ 134217727     │
│ test         │  102025 │ t            │ 134217728     │ 268435455     │
│ test         │  102026 │ t            │ 268435456     │ 402653183     │
│ test         │  102027 │ t            │ 402653184     │ 536870911     │
│ test         │  102028 │ t            │ 536870912     │ 671088639     │
│ test         │  102029 │ t            │ 671088640     │ 805306367     │
│ test         │  102030 │ t            │ 805306368     │ 939524095     │
│ test         │  102031 │ t            │ 939524096     │ 1073741823    │
│ test         │  102032 │ t            │ 1073741824    │ 1207959551    │
│ test         │  102033 │ t            │ 1207959552    │ 1342177279    │
│ test         │  102034 │ t            │ 1342177280    │ 1476395007    │
│ test         │  102035 │ t            │ 1476395008    │ 1610612735    │
│ test         │  102036 │ t            │ 1610612736    │ 1744830463    │
│ test         │  102037 │ t            │ 1744830464    │ 1879048191    │
│ test         │  102038 │ t            │ 1879048192    │ 2013265919    │
│ test         │  102039 │ t            │ 2013265920    │ 2147483647    │
└──────────────┴─────────┴──────────────┴───────────────┴───────────────┘
(32 rows)

The hash function is the native hash function defined for each type in Postgres. You can call it from SQL using the worker_hash function. You can also get the shard ID for a value directly using get_shard_id_for_distribution_column .

[Rohit25negi]

Could you please explain it a bit more? How is consistent hashing used here?

Do you mean that these Ranges are the range in the "ring" in the consistent hashing Concept?
Also, Only one hashing function is used, right? How does Citus ensure the even distribution of Data?

And, If we remove the Shard 102008(First one) if we don't need it, how will the Citus react in terms of key rearrangement? Same question for what happens when we add a new shard.

[Rohit25negi]

For the last two questions, I believe "generating multiple hashes for each node" is used to evenly distribute the data across the ring and handle shard removal and addition in a fair manner. Wanted to Check Do we have something like that in Citus?

[marcocitus]

Do you mean that these Ranges are the range in the "ring" in the consistent hashing Concept?

yes

Also, Only one hashing function is used, right? How does Citus ensure the even distribution of Data?

Shards in Citus are a logical concept, somewhat similar to vnodes in Cassandra. Individual nodes typically have multiple logical shards. Citus can rebalance based on the number of shards, the size of the shards, or a custom strategy.
https://www.citusdata.com/blog/2021/03/13/scaling-out-postgres-with-citus-open-source-shard-rebalancer/

And, If we remove the Shard 102008(First one) if we don't need it, how will the Citus react in terms of key rearrangement?

Currently there is no facility for merging shards or dropping individual shards. You would usually just move the logical shards to a different node instead.

[Rohit25negi]

I see Ok thanks.