[kvlonge]

plan, apply, load and copy now tell a scheduler whether to retry. They exit 75 (EX_TEMPFAIL) instead of 1 when every failure is one a rerun may get past. Under --json, retryable says the same thing on the run, on each table, in plan and in the {"error": …} document. retryable is true exactly when the exit code is 75. Everything else is unchanged: 0 for success, 1 for any lasting failure, 2 for usage and 130 for a forced stop.

What already worked before this change (CHI-1241 confirmation):

• sl.http paging, auth and retries: six paginators, bearer, API-key, basic and OAuth2 auth (client credentials and refresh token, with renewal on expiry or a 401), and retries on transport errors, 429 and 5xx that honour Retry-After. packages/sqlload/tests/test_http_jobs.py covers these (retries-429-and-503, the paginator cases and the OAuth renewal tests).
• load --json report: a stable document versioned by schema_version: 1, documented on the load page, with ok matching exit 0.
• Safe reruns: batched loads keep what they committed and resume from the destination's cursor (test_batched_loads.rs). Logged tables resume from their last checkpoint (test_logged_state.py). replace swaps in only at the end.

What was missing: a scheduler couldn't tell a dropped connection from a broken configuration, because both exited 1.

Classification (LoaderError::is_retryable):

• Retryable:
  • A dropped, refused or timed-out connection, to a target or to a source while planning or loading.
  • A 429, 502, 503 or 504, or a 500 with no code or a transient ClickHouse or BigQuery code, from a warehouse.
  • An sl.http job whose retries ran out on 429, 5xx or transport errors.
  • The Postgres SQLSTATEs 08000, 08003, 08006, 53200, 53300, 57P01–57P03, 40001, 40P01 and 55P03. 08P01 is excluded because PgBouncer also sends it for an unknown database.
  • A table another run is loading.
  • A Python table that raised ConnectionError or TimeoutError.
• Not retryable: configuration problems, plan errors (other than an unreachable source), credentials, permissions, data errors, 4xx responses, requests that can't be built, and answers that aren't the JSON expected.
• Object stores count only for a connection that dropped or timed out. object_store 0.14 doesn't export the error that carries a status after its own retries.
• A target failure while a source streams keeps its classification, because read_source remembers the publisher's error. CDC loads without --follow report it too.

Tests:

• Rust: a table of 22 errors classified retryable or not, and exit 75 versus 1 through the command path: a refused target, an unreachable source on its own (75) and beside a lasting plan error (1), an invalid configuration (1).
• Python, end to end:
  • A logged, checkpointed table and a cursor table both fail on the network: exit 75, retryable: true. A rerun loads every row exactly once.
  • A network failure beside a ValueError: exit 1, and each table says which kind it is.
  • An sl.http 500 after its retries: exit 75. A 404: exit 1.

Docs: an exit code table entry and a "Retrying failed runs" section in cli/index.mdx, plus the new fields in cli/load.mdx, marked "From 0.34.0 (not released yet)".

Checks: make check (fensu 0 faults), make test-unit, and make test-python (340 passed). Reviewed once, plus a follow-up check of the fixes.

[cloudflare-workers-and-pages]

Deploying sqlload with    Cloudflare Pages

Latest commit:	6098824
Status:	✅  Deploy successful!
Preview URL:	https://427ee4ae.sqlload.pages.dev
Branch Preview URL:	https://feat-retryable-exit-codes.sqlload.pages.dev

View logs