We pass a timeout into the connection pool call which should be enough

Maksim Kaitmazian authored 5 months ago and Maksim Kaitmazian committed 5 months ago

This commit resolves the performance issues observed in pgbench
benchmark*. Because of the Nagle's algorithm, every message was read with
50ms delay, while the query execution took only 1ms. By disabling the
algorithm via TCP_NODELAY option, we achieve 50 times performance
improvement for a single client in simple query mode.

* https://git.picodata.io/picodata/picodata/picodata/-/issues/1002

Denis Smirnov authored 5 months ago and Dmitry Rodionov committed 5 months ago

There was a bug when SQL statistics LRU removed some old query
from the _sql_query table. The query ref_counter field in unsigned,
but the code decremented it below zero in the space on eviction.
That produced an error (fixed).

Дмитрий Кольцов authored 5 months ago and Dmitry Rodionov committed 5 months ago

Due to the bug in RPC (see https://git.picodata.io/picodata/picodata/picodata/-/issues/991)
RPC call timeout does not account TCP connection establishment.
So a request to a peer without a connection was not aborted
due to timeout.
Force a timeout directly for a future returned by RPC to avoid this

Closes #991

Дмитрий Кольцов authored 5 months ago and Dmitry Rodionov committed 5 months ago

Previously webUI failed to retrieve data about
instances if one of those was dead. The cause was
that webUI HTTP request timeout was less than
get_instances_data timeout.

References #943

Дмитрий Кольцов authored 5 months ago and Georgy Moshkin committed 5 months ago

All migrations are explicitly listed in manifest.yaml,
so there is no much sense in strict extension check

Closes #990

Kurdakov Alexander authored 5 months ago and Alexander Kurdakov committed 5 months ago

New argument `--timeout` for `picodata expel`.

Kurdakov Alexander authored 5 months ago and Alexander Kurdakov committed 5 months ago

New argument `--timeout` for `picodata connect`.

Note that after this merge request, we no longer validate whether the format
of the created object matches with the schema. Consequently, we do not
suppress errors if formats match anymore.

Before:
```sql
CREATE USER u WITH PASSWORD 'Passw0rd'
1
CREATE USER u WITH PASSWORD 'Passw0rd'
0
CREATE USER u WITH PASSWORD 'DifferentPassw0rd'
error: user already exists with different auth method
```

Now:
```sql
CREATE USER u WITH PASSWORD 'Passw0rd'
1
CREATE USER u WITH PASSWORD 'Passw0rd'
error: user u already exists
CREATE USER u WITH PASSWORD 'DifferentPassw0rd'
error: user u already exists
CREATE USER IF NOT EXISTS u WITH PASSWORD 'DifferentPassw0rd'
0
```

See https://git.picodata.io/picodata/picodata/sbroad/-/issues/837

Up to this patch `instance_uuid` and `replicaset_uuid` were generated
using name-based UUID v3. So all instances named `i1` in all clusters
had the same UUID.

With this patch UUIDs become random v4. Particular UUID values are
generated in two locations:

1. `start_boot` (src/lib.rs)
2. `handle_join_request_and_wait` -> `build_instance` (src/rpc/join.rs)

Function `Instance::new()` is removed in favor of explicit struct
construction.

The main purpose is to reduce the usage of `Instance::new()` function.
It's going to be removed soon. No behavior changes are introduced yet.

Kurdakov Alexander authored 6 months ago and Yaroslav Dynnikov committed 6 months ago

Do not ban admin console via unix socket
after several failed attempts via `picodata connect`.