- Support groupby, aggregates and having
when the child has Global distribution:
in this case we simply don't do 2-stage
aggregation and execute everything in stage
on one node.

- subqueries reading from global tables are supported
- subqueries in queries which scan the global tables
are allowed:
`select a from global where b in (select c from segment)`

However currently using global tables in join is not
supported.

All materialized table on the storages are now fully temporary and
do not leave any records in WAL. Statistics table are left data
temporary as there is no need to recreate them on every restart.

BREAKING CHANGE!: renamed __SBROAD_STAT into _sql_stat,
__SBROAD_QIERY into _sql_query.

1. _sql_stat primary key also includes the parent_span cause
   we support PG extended protocol and we need to bind same spans
   to different trees.
2. Fix some bugs in the telemetry for queries.

Arseniy Volynets authored 1 year ago and Denis Smirnov committed 1 year ago

- now Motion(Full) has Global distribution
- Relational::Values has Any distribution
- Replicated distribution was removed, Distribution::Any is
used instead for projection with constants

feat: add checks on table columns creation and insertion, add check on parameters number, add support of is_nullable metadata

When reading from multiple storages it is
possible that some requests will execute faster
than the others. And it could lead to wrong
results when vhard moved the buckets between
nodes.

For example: we execute `select * from t`
on storage 1 and get the results. Vshard
moves the data from storage 1 to stroage 2,
and then we execute `select * from t` and in
result we get the data (that moved) twice.

This commit fixes it by using vshard's
storage ref api. Now multi-node read requests
first create the ref on each node which
blocks vshard rebalancing. Then we call
our stored procedure and remove the ref.

Maksim Kaitmazian authored 1 year ago and Denis Smirnov committed 1 year ago

Protocol has been moved to a separate repository called pgproto,
so there is no need to store it here anymore.

- add support for new dml operator update
- subqueries are not allowed in update
expression.
- primary key columns can't be updated.
- update as other dml operations is
not atomic: it may fail on some
storages and succeed on others.

Picodata was bumped to rust 1.71.0, so sbroad as its submodule can
finally update the rust version as well.