Berkay Antmen
Berkay Antmen
Are there any plans for this?
@jkatz Thank you, this is very helpful and your estimates are accurate. `SHOW shared_buffers;` gives me `5463726`. Assuming that the unit is a 8 kb page, this is 43.7 gb....
> Assuming you've performed your index build and haven't added any data to the index since @jkatz That's correct. Given the colocation, I would expect much better IO throughput. I...
Hi @ankane, I have not been testing on a snapshop but I'll keep that in mind for future benchmarks. I had previously used pg_prewarm to load the embedding table index...
@jkatz thanks for your suggestion, in here and in the email thread. 1. I will experiment with `r6gd.2xlarge` 2. If I created an HNSW index, would that not have a...
I migrated the database from `r5.2xlarge` to `r6gd.2xlarge` and performance characteristics are identical: - The query performance is entirely determined by how many blocks were a cache miss - We...
@sandys I don't think we reached a consensus in the above discussion but my understanding is there is a very significant read IO bottleneck for the IVFFLAT index implementation. I...
That's a good question. I was just in the process of setting up a database outside of AWS. I'll report back with my results comparing the same query + data....
> Is this an AWS issue? Or is this happening with any postgresql? Using the same configuration, I reproduced similar results outside of AWS. @sandys
You can do the following - have your data take less space - configure number of lists and number of probes - can the data be partitioned somehow? is there...