The pool maxwait metric currently operates differently from Pgbouncer.
The way it operates today is that we keep track of max_wait on each connected client, when SHOW POOLS query is made, we go over the connected clients and we get the max of max_wait times among clients. This means the pool maxwait will never reset, it will always be monotonically increasing until the client with the highest maxwait disconnects.
This PR changes this behavior, by keeping track of the wait_start time on each client, when a client goes into WAITING state, we record the time offset from connect_time. When we either successfully or unsuccessfully checkout a connection from the pool, we reset the wait_start time.
When SHOW POOLS query is made, we go over all connected clients and we only consider clients whose wait_start is non-zero, for clients that have non-zero wait times, we compare them and report the maximum waiting time as maxwait for the pool.
* Revert "Reset wait times when checked out successfully (#656)"
This reverts commit ec3920d60f.
* Revert "Not sure how this sneaked past CI"
This reverts commit 4c5498b915.
* Revert "only report wait times from clients currently waiting to match behavior of pgbouncer (#655)"
This reverts commit 0e8064b049.
What is wrong
Stats reported by SHOW POOLS seem to be leaking. We see lingering cl_idle , cl_waiting, and similarly for sv_idle , sv_active. We confirmed that these are reporting issues not actual lingering clients.
This behavior is readily reproducible by running
while true; do
psql "postgres://sharding_user:sharding_user@localhost:6432/sharded_db" -c "SELECT 1" > /dev/null 2>&1 &
done
Why it happens
I wasn't able to get to figure our the reason for the bug but my best guess is that we have race conditions when updating pool-level stats. So even though individual update operations are atomic, we perform a check then update sequence which is not protected by a guard.
https://github.com/postgresml/pgcat/blob/main/src/stats/pool.rs#L174-L179
I am also suspecting that using Relaxed ordering might allow this behavior (I changed all operations to use Ordering::SeqCst but still got lingering clients)
How to fix
Since SHOW POOLS/SHOW SERVER/SHOW CLIENTS all show the current state of the proxy (as opposed to SHOW STATS which show aggregate values), this PR refactors SHOW POOLS to have it construct the results directly from SHOW SERVER and SHOW CLIENT datasets. This reduces the complexity of stat updates and eliminates the need for having locks when updating pool stats as we only care about updating individual client/server states.
This will change the semantics of maxwait, so instead of it holding the maxwait time ever encountered by a client (connected or disconnected), it will only consider connected clients which should be okay given PgCat tends to hold on to client connections more than Pgbouncer.
* Refactor stats to use atomics
When we are dealing with a high number of connections, generated
stats cannot be consumed fast enough by the stats collector loop.
This makes the stats subsystem inconsistent and a log of
warning messages are thrown due to unregistered server/clients.
This change refactors the stats subsystem so it uses atomics:
- Now counters are handled using U64 atomics
- Event system is dropped and averages are calculated using a loop
every 15 seconds.
- Now, instead of snapshots being generated ever second we keep track of servers/clients
that have registered. Each pool/server/client has its own instance of the counter and
makes changes directly, instead of adding an event that gets processed later.
* Manually mplement Hash/Eq in `config::Address` ignoring stats
* Add tests for client connection counters
* Allow connecting to dockerized dev pgcat from the host
* stats: Decrease cl_idle when idle socket disconnects
