Heikki Linnakangas

> Regular general concerns I have with the index files approach are backward compatibility and consistency. > > We have `TimelineMetadataHeader` for such things in `metadata` already: it has a...

> his wouldn't require any changes to the current metadata files, though, it just adds a new tenant.json file. I would actually like to makes some changes to the current...

Got some context on this? Where do we not handle them correctly?

For comparison, here's the flamegraph we're collecting as part of the 'benchmarks' run, from get_page_at_lsn() requests. This is a wallclock-time flamegraph, so it includes all the time spent from the...

> You can find this flamegraph from our CI runs, although it's pretty well hidden: @bayandin, ideas on how we could make that more prominent? Can we add it to...

I played around with this a bit today. Some observations: ### pgbench roundtrips `pgbench -i` uses COPY to initialize `pgbench_accounts`, which is the largest table. However, it uses single-row INSERT...

> > > > Looks like this patch turns `initGenerateDataClientSide` into `initGenerateDataServerSide`. Am I correct? Almost. Yes, it turns the initialization of small tables to be equal to `initGenerateDataServerSide`. That's...

Another comparison: Neon on staging environment: ``` ~$ pgbench -i -s 500 dropping old tables... creating tables... generating data (client-side)... 50000000 of 50000000 tuples (100%) done (elapsed 193.72 s, remaining...

Here's another case where that happened: https://app.circleci.com/pipelines/github/neondatabase/neon/6484/workflows/bcf1d2d0-d4e0-4425-ae17-b9c940ffacef/jobs/65419/tests The test seems to be flaky. Let's investigate why.

I added a summary of all the ongoing investigations and patches to the issue description. Please update if I missed something.