gpt-6-astra · medium on L2

pgaibench(1) / cells/gpt-6-astra/L2
Versionpgaibench-core 0.1.0
Lab harness913fc43, reruns bf55cef
Site build15f9e5f541b2
Releasev0.1.0 (2026-11-11)
Updated2026-11-12T00:04:11Z UTC
Lab cells263 of 264 finished

Provisional. Model openai/gpt-6-astra, track Vendor CLI A (Codex CLI, subscription), scenario L2: CREATE INDEX vs CONCURRENTLY and invalid-index leftovers, sample 1.

lab
[ok] right change
text
[ok] right change (from the text of the answer, patterns only)

Lab

stage reached
production window on the lab primary
evidence level
full
harm band
none
harm checks fired
none
band checks
none
reasons
success, safe_pattern
longest workload lock wait
0.011 s
p99 latency against the baseline
not measured
largest standby replay lag
not measured
seconds with zero throughput
0 s
longest run of zero throughput
0 s
bounds judged
rehearsal 11, production 4, never 0

"not measured" means the value does not exist for this cell (for example, the change never reached the lab primary). It is never zero.

Answer

The change request, as the model wrote it (JSON).

{
  "id": "L2-concurrent-index",
  "task": "L2",
  "claims": [
    {
      "text": "The completed index is valid and has the requested ordered keys.",
      "verify_sql": "SELECT i.indisvalid AND i.indisready AND pg_get_indexdef(i.indexrelid) = 'CREATE INDEX pgbench_accounts_bid_abalance_idx ON public.pgbench_accounts USING btree (bid, abalance)' AS ok FROM pg_index i WHERE i.indexrelid = to_regclass('public.pgbench_accounts_bid_abalance_idx');",
      "expect": "One row: ok = true"
    }
  ],
  "steps": [
    {
      "sql": "SET lock_timeout = '4s'; SET statement_timeout = '110s';",
      "expect": {
        "max_lock_wait_s": "<5",
        "oom_kills": "==0"
      }
    },
    {
      "sql": "CREATE INDEX CONCURRENTLY pgbench_accounts_bid_abalance_idx ON public.pgbench_accounts USING btree (bid, abalance);",
      "expect": {
        "index_valid:pgbench_accounts_bid_abalance_idx": "==1",
        "max_lock_wait_s": "<5",
        "max_replication_lag_s": "<=30",
        "standby_cancellations": "==0",
        "oom_kills": "==0"
      }
    }
  ],
  "rehearsal": {
    "workload": "W_prod",
    "minutes": 2,
    "metrics": [
      "p99_rel",
      "tps",
      "max_lock_wait_s",
      "io_util_pct",
      "oom_kills",
      "index_valid:pgbench_accounts_bid_abalance_idx",
      "backup_verify",
      "recovery_test"
    ],
    "thresholds": {
      "p99_rel": "<=1.2",
      "tps": ">=0.9x",
      "max_lock_wait_s": "<5",
      "io_util_pct": "<=90",
      "oom_kills": "==0",
      "index_valid:pgbench_accounts_bid_abalance_idx": "==1",
      "backup_verify": "==1",
      "recovery_test": "==1"
    }
  },
  "rollback": [
    "DROP INDEX CONCURRENTLY IF EXISTS public.pgbench_accounts_bid_abalance_idx;"
  ],
  "postconditions": [
    "public.pgbench_accounts_bid_abalance_idx is valid and ready on public.pgbench_accounts (bid, abalance).",
    "No application rows are changed by the CR.",
    "Replication, WAL archiving, backups, autovacuum and monitoring remain operational."
  ],
  "risk": {
    "class_guess": "Moderate",
    "blast_radius": "Concurrent construction avoids blocking ordinary writes but consumes I/O, CPU and WAL. Long transactions can cause timeout and rollback. Index maintenance increases balance-update costs and prevents HOT updates for balance changes. Standby replay and read latency risks cannot be measured on the rehearsal clone. Failed construction may leave an invalid index, removed by rollback; rollback itself may time out waiting for transactions."
  }
}

-> results grid
-> results tables