Arseniy Volynets authored 8 months ago and Dmitry Rodionov committed 8 months ago

- Some sql queries that read from sharded tables
can fail with timeout error, but because of rethrows
it is showed as 'Connection reset error'. In order
to avoid retrying such tests in CI we can use retries
in tests. Note: solution to the problem itself is in
the progress, but we don't know when it will be done:
https://github.com/tarantool/vshard/pull/442

- Another solution is to use global tables instead of
sharded tables in tests where it does not matter.

- This commit uses both approaches: when possible sharded
tables were replaced with global ones. In other cases retries
were used.

- when doing global dml sql query, there is error
like user don't have access to 'raft_state` space.
Fix this by doing parts of pipeline that access
system spaces under admin

This is needed because we require Service::on_start callbacks to be
idempotent because governor may end up call them multiple times.

Use "/path" RPC routes to call plugin-defined endpoints.
Use ".proc" RPC routes to call builtin picodata endpoints (for example
.proc_instance_info)

- use RouteBuilder to register rpc endpoints (input and ouptut is raw
  bytes for now)
- use RequestBuilder to send RPC requests to other instances which are
  running a specified plugin.service (guarded by version)
    - supported routing by
        - explicit instance_id
        - replicaset_id
        - bucket_id
        - any available instance with given service
- use context::Context to extract information about request or add
  custom fields at runtime (middleware not yet implemented)

The problem is MsgTimeoutNow always results in a new election, because
it explicitly ignores the pre_vote step. This results in integration
test flakiness whenever a follower calls promote_or_fail while the
leader's raft log is longer.

We've cherry-picked some fixes but didnt cut a release yet. Prepare
changelog.