fix: retry refresh token updates on serialization failures (SQL & ent)#4830
fix: retry refresh token updates on serialization failures (SQL & ent)#4830RonnanSouza wants to merge 1 commit into
Conversation
RonnanSouza
force-pushed
the
fix/sql-retry-serialization-failure
branch
3 times, most recently
from
cf70e72 to
1f86e6a
Compare
UpdateRefreshToken runs a read-modify-write inside a SERIALIZABLE transaction. Under concurrent refresh-token rotation of the same token (e.g. multiple kube clients refreshing at once), Postgres aborts the conflicting transaction with SQLSTATE 40001 and the error surfaced to clients as HTTP 500. Add a bounded, jittered retry around the refresh-token update that re-runs the transaction on transient serialization/deadlock failures, detected per-driver (Postgres 40001/40P01, MySQL 1213/1205). Error wraps in the SQL storage now use %w so the driver error can be matched with errors.As. SQLite serializes writes and opts out (nil check). Enables the previously-disabled refresh-token concurrency conformance tests for Postgres and MySQL. Signed-off-by: Ronan <ronanpalmeiras@gmail.com> fix(storage/ent): retry refresh token update on serialization failures Mirror the SQL backend fix in the ent storage. UpdateRefreshToken runs in a SERIALIZABLE transaction and aborts with a serialization failure under concurrent rotation of the same token, surfacing as HTTP 500. Wrap the refresh-token update in a bounded, jittered retry that re-runs the transaction on transient serialization/deadlock failures, detected per-driver (Postgres 40001/40P01, MySQL 1213/1205) via errors.As over the ent-wrapped driver error. Enables the previously-disabled refresh-token concurrency conformance tests for Postgres, MySQL and MySQL 8. Signed-off-by: Ronan <ronanpalmeiras@gmail.com> refactor(storage): extract shared SQL retry policy into sqlretry refactor(storage): extract shared SQL retry policy into sqlretry Signed-off-by: Ronan <ronanpalmeiras@gmail.com>
RonnanSouza
force-pushed
the
fix/sql-retry-serialization-failure
branch
from
1f86e6a to
a016e56
Compare
| if err != nil { | ||
| return err | ||
| } | ||
| if r, err = updater(r); err != nil { |
There was a problem hiding this comment.
Running updater(r) inside a retry can be unsafe in many scenarios... The least I would require for a feature like this would be to make it opt-in
There was a problem hiding this comment.
Running
updater(r)inside a retry can be unsafe in many scenarios... The least I would require for a feature like this would be to make it opt-in
I don't see how this can be unsafe, if the it re-opens a fresh transaction and re-reads state each attempt. Also, I couldn't think of a scenario where a straight 500 is better than a retry with a reasonable timeout. But if you folks thinks this is necessary to have it merged, I'm happy to adapt
There was a problem hiding this comment.
Running updater(r) multiple times can have all sorts of consequences depending on the connector being used. This introduces a large pool of failure modes.
-
For OIDC connectors backed by providers with refresh-token rotation, the first attempt can successfully exchange the upstream refresh token and receive a new one, then lose the SQL transaction. The
retry can call the provider again with the old upstream refresh token. Providers like Okta may reject that as reuse, and may invalidate the newly issued token chain. -
The updater mutates captured state outside the transaction (
newToken,lastUsed,ident,rerr). If attempt 1 reaches the connector and attempt 2 takes a different branch after re-reading the DB
row, the final response/offline-session update can be based on state produced by a failed attempt. -
Even when the upstream refresh call is technically reusable, retrying it can duplicate side effects: extra token endpoint calls, rate-limit pressure, duplicate audit/security events, and slower refresh
responses under contention.
I'd strongly advise not turning on this code by default. In Flux, when we change default functionality that can have a wide impact like this, we ship it under a feature flag. If we don't hear back about problems in that feature and we think it's a good behavior to have by default, we flip the flag's default value later after a few quarters. I think this is what a GA CNCF project with 10k stars like Dex should do for a feature with this impact surface.
|
I'd also consider making the hard-coded parameters inputs with defaults (i.e. the number of retries, backoff, jitter). Maybe the hard-coded / defaults are good for some users, but others may need to fine-tune... |
|
Hey Pimenta, thanks for the review.
I set this way because I saw a similar approach in the k8s retry mechanism. But I can make it customizable with the enable/disable config from the previous comment. |
2 participants
Overview
When several clients call
UpdateRefreshToken— which performs a read-modify-write inside a SERIALIZABLE transaction — concurrent rotation of the same token makes the database abort one transaction with a serialization failure, e.g.:
time=2026-06-10T09:34:44.629Z level=ERROR msg="failed to update refresh token" err="update refresh token: pq: could not serialize access due to concurrent update" Neither SQL storage backend retried, so the abort surfaced to clients as a 500.
What this PR does / why we need it
Add a bounded, jittered retry around the refresh-token update in both SQL backends (
storage/sqlandstorage/ent):errors.As: Postgres40001/40P01, MySQL1213/1205.database/sqlnor ent retries transactions natively.Special notes for your reviewer