Data error catalogue

1. The table

Reason	HTTP	Means	Fix
accessDenied	403	insufficient permissions for the resource	guides/identity/fix-permission-denied.md — and note BigQuery permissions are often dataset-level
quotaExceeded	403	the project exceeded a BigQuery quota	read the live quota; back off; request an increase if adjustable
rateLimitExceeded	403, 429	too many requests, too quickly	retryable — exponential backoff
notFound	404	the referenced resource does not exist	check the fully-qualified name, including the project
billingNotEnabled	403	billing is not enabled on the project	not a BigQuery problem at all
billingTierLimitExceeded	400	an on-demand query exceeded tier 100	see §3
invalidQuery	400	syntax	fix the SQL
resourcesExceeded	400	the job "consumed excessive resources"	see §4
badRequest	400	invalid input, or a streaming buffer conflict	see §5
duplicate	409	the resource already exists	make the create idempotent
backendError	500, 502, 503, 504	service-side	retryable — exponential backoff
internalError	500	service-side	retryable
timeout	400	the job exceeded a time limit	split the job, or raise the limit where one exists
tableUnavailable	400	a third-party table is temporarily unavailable	retry later; this is an external dependency, not your query

2. The retryable set, as the vendor lists it

backendError, internalError, jobBackendError, jobInternalError, rateLimitExceeded — with exponential backoff.

Everything else in §1 is a change, not a retry. A client that retries invalidQuery or billingTierLimitExceeded will do so forever.

3. billingTierLimitExceeded is the cost model talking

It means an on-demand query exceeded tier 100. So it is not a bug in the query and not a quota you ask to raise: it is the on-demand model refusing a query of that size.

Three real answers:

1. reduce what the query scans — partitioning, clustering, fewer columns;
2. move the workload to capacity-based pricing, where compute is reserved slots rather than per-query

(reference/data/bigquery.md §2);

1. split the job.

This row is why reference/data/bigquery.md insists the pricing model is a design decision. It is the only error in the catalogue whose fix may be "change how you buy compute".

4. resourcesExceeded is about the shape of the query

The job "consumed excessive resources". In practice that is a query doing something that will not distribute — an enormous ORDER BY over the whole result, a window function without partitioning, an unbounded self-join.

Check whether the query has a stage that must be processed on one worker. Fix by removing it, not by buying more slots: this failure is not usually a capacity shortage.

5. badRequest has two very different causes

"Invalid input or streaming buffer conflict". The second one is the non-obvious half: rows recently streamed into a table sit in a buffer, and some operations against a table with a non-empty streaming buffer are refused.

Check whether the table has had recent streaming inserts before concluding the request itself was malformed.

6. Dataflow and Dataproc

No vendor error table was read for either in this release, so none is invented. Two facts from their own pages carry the most diagnostic weight:

• Dataflow is exactly-once by default, with at-least-once available — so duplicate output from a Dataflow

pipeline is a configuration question first (reference/data/dataflow.md §3);

• a Dataproc cluster create, scale or shutdown takes "90 seconds or less, on average" — so an operation

hanging for much longer than that is not normal slowness (reference/data/dataproc.md §4).

Recorded as a gap in changelog/2026-10-21-data-family.md.