What happened?
Copied from my discord msg where a member of the pgai team suggested I create a gh issue (I didn’t know if this was appropriate here):
My index has been building for 18 days!!
I'm just wondering if something is misconfigured for the next time I build an index of this magnitude. I have 134 million 1024-dim vectors with the following setup:
System specs:
- 64gb ram
- 6 core EPYC processor
- 2x nvme ssd (not striped in a RAID setup, I kept one for the os + index and the other for the working set since it is 500gb on disk)
Postgres configuration:
- 3 parallel workers
- 16gb maintenance_work_mem
- 8gb shared buffers
vmstat shows no disk swapping
htop shows each worker slowly filled up over a couple days to reach a maximum of 10gb memory per worker process. It has yet to go higher. Thus, system memory has not exceeded 30gb of ram as htop displays.
Also, CPU usage has slowly dropped down to the ~17-20% it sits at now.
I feel as though I'm close to the end of this index build but because of the current bug with parallel index building not reporting to pg_stat properly, I have little idea besides checking the pg_indexes folder. I've attached a screenshot of my tmux panes.
Any ideas from anyone?
pgvectorscale extension affected
0.9.0
PostgreSQL version used
18
What operating system did you use?
Ubuntu 21.04
What installation method did you use?
Source
What platform did you run on?
On prem/Self-hosted
Relevant log output and stack trace
How can we reproduce the bug?
Try to create vector embeddings for the entire USPTO dataset lol
Are you going to work on the bugfix?
None
What happened?
Copied from my discord msg where a member of the pgai team suggested I create a gh issue (I didn’t know if this was appropriate here):
My index has been building for 18 days!!
I'm just wondering if something is misconfigured for the next time I build an index of this magnitude. I have 134 million 1024-dim vectors with the following setup:
System specs:
Postgres configuration:
vmstat shows no disk swapping
htop shows each worker slowly filled up over a couple days to reach a maximum of 10gb memory per worker process. It has yet to go higher. Thus, system memory has not exceeded 30gb of ram as htop displays.
Also, CPU usage has slowly dropped down to the ~17-20% it sits at now.
I feel as though I'm close to the end of this index build but because of the current bug with parallel index building not reporting to pg_stat properly, I have little idea besides checking the pg_indexes folder. I've attached a screenshot of my tmux panes.
Any ideas from anyone?
pgvectorscale extension affected
0.9.0
PostgreSQL version used
18
What operating system did you use?
Ubuntu 21.04
What installation method did you use?
Source
What platform did you run on?
On prem/Self-hosted
Relevant log output and stack trace
How can we reproduce the bug?
Try to create vector embeddings for the entire USPTO dataset lolAre you going to work on the bugfix?
None