prole/demo/ci-migration/07_corrected_functions.sql

-- 07_corrected_functions.sql
-- Corrected versions of all four utility functions from the change request.
-- Each fix is annotated with the issue number it addresses (see REVIEW.md).

-- ── Fix for cleanup_expired_artifacts ────────────────────────────────────────
-- Issue 4: Added partition_key to WHERE and IN-list for partition pruning.
-- Issue 4: Replaced FOR UPDATE with FOR UPDATE SKIP LOCKED to prevent
--          stacking behind a concurrent cleanup job.
-- Bonus: RETURNING clause allows callers to log what was actually deleted.
CREATE OR REPLACE FUNCTION demo.cleanup_expired_artifacts(
    p_batch_size INT DEFAULT 10000
)
RETURNS INT AS $$
DECLARE
    v_deleted INT;
BEGIN
    DELETE FROM demo.ci_job_artifacts
    WHERE (id, partition_key) IN (
        SELECT id, partition_key
        FROM   demo.ci_job_artifacts
        WHERE  expire_at      < NOW()
          AND  locked         = FALSE
          AND  partition_key  < DATE_TRUNC('month', NOW())::DATE  -- prune to old partitions only
        ORDER BY expire_at ASC
        LIMIT  p_batch_size
        FOR UPDATE SKIP LOCKED
    );

    GET DIAGNOSTICS v_deleted = ROW_COUNT;
    RAISE NOTICE 'Cleaned up % expired artifacts', v_deleted;
    RETURN v_deleted;
END;
$$ LANGUAGE plpgsql;


-- ── Fix for terminate_idle_sessions ──────────────────────────────────────────
-- Issue 5: Added filter to exclude PgBouncer pooler connections.
--          In transaction-mode PgBouncer the backend appears idle between
--          client transactions. Terminating it breaks the pool and causes
--          "SSL connection has been closed unexpectedly" for active clients.
-- Bonus: filter also excludes replication and background workers.
CREATE OR REPLACE FUNCTION demo.terminate_idle_sessions(
    p_idle_threshold INTERVAL  DEFAULT '10 minutes',
    p_max_terminated INT       DEFAULT 50,
    p_pgbouncer_app  TEXT      DEFAULT 'pgbouncer'   -- set to your app_name
)
RETURNS TABLE(pid INT, username TEXT, idle_duration INTERVAL, query TEXT) AS $$
BEGIN
    RETURN QUERY
    WITH idle_sessions AS (
        SELECT
            a.pid,
            a.usename::TEXT,
            NOW() - a.state_change          AS duration,
            a.query
        FROM pg_stat_activity a
        WHERE a.state          = 'idle'
          AND NOW() - a.state_change > p_idle_threshold
          AND a.pid            <> pg_backend_pid()
          AND a.backend_type   = 'client backend'
          AND a.application_name NOT ILIKE '%' || p_pgbouncer_app || '%'
          AND a.client_addr IS NOT NULL           -- exclude unix socket (local tools)
        ORDER BY duration DESC
        LIMIT p_max_terminated
    )
    SELECT s.pid, s.usename, s.duration, s.query
    FROM idle_sessions s
    WHERE pg_terminate_backend(s.pid);
END;
$$ LANGUAGE plpgsql;


-- ── Fix for calculate_namespace_storage ──────────────────────────────────────
-- Issue 6: Cache race condition fixed with FOR UPDATE SKIP LOCKED.
--          Callers that can't acquire the cache row lock return the stale
--          cached value rather than all recomputing simultaneously.
-- Issue 7: Recursive CTE depth limit enforced (max 20 levels).
-- Requires tables: namespaces(id, parent_id), project_storages(namespace_id, storage_bytes),
--                  namespace_storage_cache(namespace_id, total_bytes, calculated_at)
-- (These don't exist in our demo schema; this is illustrative code only.)
CREATE OR REPLACE FUNCTION demo.calculate_namespace_storage(
    p_namespace_id BIGINT
)
RETURNS BIGINT AS $$
DECLARE
    v_total_bytes   BIGINT := 0;
    v_cached_bytes  BIGINT;
    v_cached_at     TIMESTAMP;
BEGIN
    -- Attempt to lock the cache row exclusively.
    -- SKIP LOCKED means concurrent callers return the stale value immediately
    -- instead of piling up and all recomputing the expensive CTE.
    SELECT total_bytes, calculated_at
    INTO   v_cached_bytes, v_cached_at
    FROM   namespace_storage_cache
    WHERE  namespace_id = p_namespace_id
    FOR UPDATE SKIP LOCKED;

    -- Cache hit: return without recomputing
    IF v_cached_at IS NOT NULL AND v_cached_at > NOW() - INTERVAL '1 hour' THEN
        RETURN v_cached_bytes;
    END IF;

    -- Cache miss or expired: recompute with depth-limited recursive CTE.
    -- Issue 7 fix: depth column added to prevent infinite recursion on cycles.
    WITH RECURSIVE ns_tree AS (
        SELECT id, 0 AS depth
        FROM   namespaces
        WHERE  id = p_namespace_id

        UNION ALL

        SELECT n.id, t.depth + 1
        FROM   namespaces n
        JOIN   ns_tree    t ON n.parent_id = t.id
        WHERE  t.depth < 20           -- hard cap; cyclic namespaces stop here
    )
    SELECT COALESCE(SUM(ps.storage_bytes), 0)
    INTO   v_total_bytes
    FROM   ns_tree        nt
    JOIN   project_storages ps ON ps.namespace_id = nt.id;

    INSERT INTO namespace_storage_cache (namespace_id, total_bytes, calculated_at)
    VALUES (p_namespace_id, v_total_bytes, NOW())
    ON CONFLICT (namespace_id) DO UPDATE
        SET total_bytes    = EXCLUDED.total_bytes,
            calculated_at  = EXCLUDED.calculated_at;

    RETURN v_total_bytes;
END;
$$ LANGUAGE plpgsql;


-- ── Fix for database_health view ─────────────────────────────────────────────
-- Issue 8: NULLIF prevents division by zero if n_live_tup is ever 0.
-- The original WHERE n_live_tup > 1000 makes the crash unlikely in practice,
-- but this is fragile to future refactoring that removes the filter.
CREATE OR REPLACE VIEW demo.database_health AS
SELECT
    schemaname,
    relname                                                                 AS table_name,
    n_live_tup                                                              AS live_rows,
    n_dead_tup                                                              AS dead_rows,
    ROUND(n_dead_tup::NUMERIC / NULLIF(n_live_tup, 0) * 100, 2)           AS dead_row_pct,
    last_vacuum,
    last_autovacuum,
    last_analyze,
    pg_size_pretty(pg_total_relation_size(schemaname || '.' || relname))   AS total_size
FROM pg_stat_user_tables
WHERE n_live_tup > 1000
ORDER BY n_dead_tup DESC;