nix: add new loadtest for errors

[steve-chavez]

Adds a new loadtest that includes only requests that produces errors (misspelled relations, misspelled functions, permission errors).

Tries to repro #4462

[steve-chavez]

So far this proves that the rate of error processing is much lower than normal successful requests.
Rate of errors is ~67 vs and mixed is 468, this while using much more CPU, ~100% on errors vs ~20% on mixed.
See https://github.com/PostgREST/postgrest/actions/runs/20441827293?pr=4576#summary-58736347815.

OOMs of #4462 are not reproduced, this looks only dependent of the number of tables, not the number of continued requests.

[steve-chavez]

Later on we can add more types of errors for this load test: https://docs.postgrest.org/en/v14/references/errors.html#postgrest-error-codes. We should ensure error processing stays fast for every case.

[steve-chavez]

This CI job proves there's considerable memory consumed when the table/function count is 30K. Unfortunately that number is causing a CI coverage failure which I don't understand. So I'll leave the table/function count at 950.

[steve-chavez]

Unfortunately that number is causing a CI coverage failure which I don't understand.

And looks like this coverage failure persist even after reducing the table count. The error reported on CI doesn't reveal anything and locally it cannot be reproduced.

[steve-chavez]

Added 20K tables now and left adding many functions as TODO (this was somehow making the loadtest slow). This now proves the memory consumption is high for the fuzzy search, I've also added a sleep for 5 seconds so it ensures the high memory consumption doesn't come from the schema cache.

[wolfgangwalther]

Looks like this argument has a default of 0, so the condition will always be true. Should work to just call sleep all the time - with 0 as argument if the default is used.

[steve-chavez]

Fixed in 76e0e1f, just removed the default instead.