sase.tr/docs/analytics-queries.sql

-- Ad-hoc analytics queries against query_log_insights.
-- All queries default to the last 30 days; adjust the interval as needed.

-- ─── 1. Cache hit ratio (db / redis_positive / redis_negative / lock_wait / miss) ───
SELECT
  cache_source,
  count(*)                                                            AS queries,
  round(100.0 * count(*) / sum(count(*)) OVER (), 1)                  AS pct,
  round(avg(response_time_ms))                                        AS avg_ms,
  percentile_cont(0.95) WITHIN GROUP (ORDER BY response_time_ms)::int AS p95_ms
FROM query_log_insights
WHERE created_at > now() - interval '30 days'
GROUP BY cache_source
ORDER BY queries DESC;

-- ─── 2. Per-source latency & success ─────────────────────────────────────────────
SELECT
  source,
  count(*)                                                            AS queries,
  round(100.0 * sum((success)::int) / count(*), 1)                    AS success_pct,
  round(avg(response_time_ms))                                        AS avg_ms,
  percentile_cont(0.50) WITHIN GROUP (ORDER BY response_time_ms)::int AS p50_ms,
  percentile_cont(0.95) WITHIN GROUP (ORDER BY response_time_ms)::int AS p95_ms,
  percentile_cont(0.99) WITHIN GROUP (ORDER BY response_time_ms)::int AS p99_ms,
  max(response_time_ms)                                               AS max_ms
FROM query_log_insights
WHERE created_at > now() - interval '30 days'
  AND cache_source = 'miss'  -- exclude cache hits for a fair upstream comparison
GROUP BY source
ORDER BY queries DESC;

-- ─── 3. Slowest WMIs (which brand prefixes hurt us most) ─────────────────────────
SELECT
  wmi,
  count(*)                                                            AS queries,
  round(avg(response_time_ms))                                        AS avg_ms,
  percentile_cont(0.95) WITHIN GROUP (ORDER BY response_time_ms)::int AS p95_ms,
  round(100.0 * sum((NOT success)::int) / count(*), 1)                AS fail_pct
FROM query_log_insights
WHERE created_at > now() - interval '30 days'
  AND cache_source = 'miss'
GROUP BY wmi
HAVING count(*) >= 3
ORDER BY p95_ms DESC
LIMIT 20;

-- ─── 4. Per-stage breakdown when an upstream is to blame ─────────────────────────
SELECT
  source,
  count(*)                                                AS queries,
  round(avg(pcat_ms))     AS avg_pcat_ms,
  round(avg(emex_ms))     AS avg_emex_ms,
  round(avg(pl24_ms))     AS avg_pl24_ms,
  round(avg(vin_api_ms))  AS avg_vin_api_ms,
  round(avg(lock_wait_ms)) AS avg_lock_wait_ms
FROM query_log_insights
WHERE created_at > now() - interval '7 days'
  AND cache_source = 'miss'
GROUP BY source
ORDER BY queries DESC;

-- ─── 5. Circuit breaker / abort frequency ────────────────────────────────────────
SELECT
  date_trunc('hour', created_at)              AS hour,
  count(*) FILTER (WHERE pl24_circuit_open)   AS cb_open_when_called,
  count(*) FILTER (WHERE pl24_skipped)        AS pl24_skipped_count,
  count(*) FILTER (WHERE aborted)             AS aborted_count,
  count(*)                                    AS total
FROM query_log_insights
WHERE created_at > now() - interval '24 hours'
GROUP BY hour
ORDER BY hour DESC
LIMIT 24;

-- ─── 6. Candidate-modal rate (how often does the user have to disambiguate?) ─────
SELECT
  result_kind,
  count(*)                                            AS queries,
  round(100.0 * count(*) / sum(count(*)) OVER (), 1)  AS pct
FROM query_log_insights
WHERE created_at > now() - interval '30 days'
GROUP BY result_kind
ORDER BY queries DESC;

-- ─── 7. Top failing VINs (cluster of retries / unsupported VINs) ─────────────────
SELECT
  vin,
  wmi,
  count(*)                          AS attempts,
  count(*) FILTER (WHERE success)   AS ok,
  count(*) FILTER (WHERE NOT success) AS fail,
  max(response_time_ms)             AS worst_ms,
  max(created_at)                   AS last_seen
FROM query_log_insights
WHERE created_at > now() - interval '30 days'
GROUP BY vin, wmi
HAVING count(*) FILTER (WHERE NOT success) >= 2
ORDER BY fail DESC, attempts DESC
LIMIT 30;

-- ─── 8. Dedup effectiveness (lock_wait shows how many requests piggy-backed) ─────
SELECT
  date_trunc('day', created_at) AS day,
  count(*) FILTER (WHERE cache_source = 'lock_wait')      AS dedup_savings,
  count(*) FILTER (WHERE cache_source = 'miss')           AS actual_decodes,
  count(*) FILTER (WHERE cache_source = 'redis_positive') AS redis_hits,
  count(*) FILTER (WHERE cache_source = 'redis_negative') AS redis_neg_hits,
  count(*) FILTER (WHERE cache_source = 'db_hit')         AS db_hits,
  count(*)                                                AS total
FROM query_log_insights
WHERE created_at > now() - interval '14 days'
GROUP BY day
ORDER BY day DESC;