fix(objects): use live PG row estimates and refresh stale stats

This commit is contained in:
zipg authored and GitHub committed 2026-09-28 21:11:00 +08:00
1 parent 708d45cea9
commit 52b7462ab2
3 files changed
+219 -13

No files matched your search

+56 -13
View File
@@ -7343,22 +7343,47 @@ pub async fn get_custom_type_details(pool: &Pool, schema: &str, name: &str) -> R
})
}
/// Row/size estimates for the object browser, per schema.
///
/// `pg_class.reltuples` only moves when ANALYZE (or autovacuum) rewrites it, so a
/// table that was written to — or never analyzed at all — keeps reporting its
/// last known count, often `0` (#10461). `pg_stat_user_tables.n_live_tup` is the
/// statistics collector's live estimate: it tracks DML within seconds and still
/// avoids a `COUNT(*)` scan, which is what the UI promises in its column hint.
const POSTGRES_OBJECT_STATISTICS_SQL: &str = "SELECT c.relname, \
GREATEST(COALESCE(s.n_live_tup, c.reltuples), 0)::bigint AS estimated_rows, \
pg_catalog.pg_total_relation_size(c.oid)::bigint AS total_bytes \
FROM pg_catalog.pg_class c \
JOIN pg_catalog.pg_namespace n ON n.oid = c.relnamespace \
LEFT JOIN pg_catalog.pg_stat_user_tables s ON s.relid = c.oid \
WHERE n.nspname = $1 AND c.relkind IN ('r','m','f','p') \
ORDER BY c.relname";
/// Some Postgres-compatible engines do not expose `pg_stat_user_tables`; fall
/// back to the ANALYZE-time estimate rather than dropping the columns entirely.
const POSTGRES_OBJECT_STATISTICS_FALLBACK_SQL: &str = "SELECT c.relname, \
GREATEST(c.reltuples, 0)::bigint AS estimated_rows, \
pg_catalog.pg_total_relation_size(c.oid)::bigint AS total_bytes \
FROM pg_catalog.pg_class c \
JOIN pg_catalog.pg_namespace n ON n.oid = c.relnamespace \
WHERE n.nspname = $1 AND c.relkind IN ('r','m','f','p') \
ORDER BY c.relname";
pub async fn list_object_statistics(pool: &Pool, schema: &str) -> Result<Vec<ObjectStatistics>, String> {
let schema = if schema.is_empty() { "public" } else { schema };
let client = checkout_postgres_client(pool, None, super::connection_timeout()).await?;
let rows = postgres_query_cached(
&client,
"SELECT c.relname, \
GREATEST(c.reltuples, 0)::bigint AS estimated_rows, \
pg_catalog.pg_total_relation_size(c.oid)::bigint AS total_bytes \
FROM pg_catalog.pg_class c \
JOIN pg_catalog.pg_namespace n ON n.oid = c.relnamespace \
WHERE n.nspname = $1 AND c.relkind IN ('r','m','f','p') \
ORDER BY c.relname",
&[&schema],
)
.await
.map_err(|e| e.to_string())?;
let rows = match postgres_query_cached(&client, POSTGRES_OBJECT_STATISTICS_SQL, &[&schema]).await {
Ok(rows) => rows,
Err(error) => {
log::warn!(
"[postgres][object-statistics] live tuple estimate unavailable, falling back to reltuples: {}",
pg_error_to_string(error)
);
postgres_query_cached(&client, POSTGRES_OBJECT_STATISTICS_FALLBACK_SQL, &[&schema])
.await
.map_err(|e| e.to_string())?
}
};
Ok(rows
.iter()
.map(|row| ObjectStatistics {
@@ -15594,4 +15619,22 @@ mod tests {
// `IN` must not match the start of `INTO`.
assert_eq!(parse_pg_partition_bound("FOR VALUES INTO (1)"), None);
}
#[test]
fn object_statistics_prefers_live_tuples_and_keeps_a_reltuples_fallback() {
// #10461: `reltuples` lags behind DML, so the live estimate must win when
// `pg_stat_user_tables` is available...
assert!(POSTGRES_OBJECT_STATISTICS_SQL.contains("pg_catalog.pg_stat_user_tables"));
assert!(POSTGRES_OBJECT_STATISTICS_SQL.contains("COALESCE(s.n_live_tup, c.reltuples)"));
// ...while an engine without that view still reports counts.
assert!(!POSTGRES_OBJECT_STATISTICS_FALLBACK_SQL.contains("pg_stat_user_tables"));
assert!(POSTGRES_OBJECT_STATISTICS_FALLBACK_SQL.contains("GREATEST(c.reltuples, 0)"));
for sql in [POSTGRES_OBJECT_STATISTICS_SQL, POSTGRES_OBJECT_STATISTICS_FALLBACK_SQL] {
// Both queries must keep the same projection so the row mapping stays valid.
assert!(sql.contains("AS estimated_rows"));
assert!(sql.contains("AS total_bytes"));
assert!(sql.contains("c.relkind IN ('r','m','f','p')"));
assert!(sql.contains("WHERE n.nspname = $1"));
}
}
}