fix: limit pg_stat_statements to top 50 queries by exec time

Instead of removing pg_stat_statements from the Prometheus sink entirely,
keep the collection but cap cardinality at 50 queryids (ranked by total
exec time). This gives 50 × 16 = 800 series max instead of 30K+.

The previous weak filter (calls >= 5 OR exec_time >= 1000) let through
thousands of queries on busy databases. ORDER BY + LIMIT 50 is a hard cap.

https://claude.ai/code/session_01SzJxzZNQjDQphaHyaX3RU7

Author: claude

config/pgwatch-prometheus/metrics.yml

@@ -1014,6 +1014,8 @@ metrics:
             join pg_database on pg_database.oid = pg_stat_statements.dbid
             where pg_database.datname = current_database()
             group by pg_database.datname, pg_stat_statements.queryid
+            order by sum(pg_stat_statements.total_exec_time) desc
+            limit 50
         )
         select
           datname as tag_datname,
@@ -1035,7 +1037,6 @@ metrics:
           temp_bytes_read::int8 as temp_bytes_read,
           temp_bytes_written::int8 as temp_bytes_written
           from aggregated_statements
-          where calls >= 5 or exec_time_total >= 1000
       17: |
         with aggregated_statements as ( /* pgwatch_generated */
           select
@@ -1061,6 +1062,8 @@ metrics:
             join pg_database on pg_database.oid = pg_stat_statements.dbid
             where pg_database.datname = current_database()
             group by pg_database.datname, pg_stat_statements.queryid
+            order by sum(pg_stat_statements.total_exec_time) desc
+            limit 50
         )
         select
           datname as tag_datname,
@@ -1082,7 +1085,6 @@ metrics:
           temp_bytes_read::int8 as temp_bytes_read,
           temp_bytes_written::int8 as temp_bytes_written
           from aggregated_statements
-          where calls >= 5 or exec_time_total >= 1000
     gauges:
       - calls
       - plans_total
@@ -2532,8 +2534,7 @@ presets:
       pg_stat_all_tables: 30
       pg_class: 30
       pg_stat_all_indexes: 30
-      # pg_stat_statements removed: handled by flask-pgss-api via Postgres sink
-      # Collecting per-queryid metrics here caused 30K+ series cardinality explosion
+      pg_stat_statements: 30
       pg_stat_replication: 30
       pg_stat_slru: 30
       pg_statio_all_tables: 30

docker-compose.yml

@@ -2,9 +2,8 @@
 # Total allocation: ~5.1 vCPUs, ~7.6 GiB RAM (with headroom for host OS).
 # CPU overcommitment is intentional - containers rarely hit limits simultaneously.
 #
-# pgwatch-prometheus gets higher CPU (1.5) to handle metric collection load.
-# Note: pg_stat_statements is NOT collected here (handled by flask-pgss-api
-# via the Postgres sink) to avoid cardinality explosion in VictoriaMetrics.
+# pgwatch-prometheus collects metrics including pg_stat_statements (top 50 queries
+# by exec time to control cardinality in VictoriaMetrics).
 
 services:
   # Config Initializer - Copies static configs from image to volume