Skip to content

fix(session): merge a save into the stored session instead of overwriting it - #189

Merged
tina4stack merged 3 commits into
tina4stack:v3from
MichaelC8E:fix/session-save-keeps-a-concurrent-logout
Oct 1, 2026
Merged

tina4stack merged 3 commits into
tina4stack:v3from
MichaelC8E:fix/session-save-keeps-a-concurrent-logout

Conversation

@MichaelC8E

@MichaelC8E MichaelC8E commented Oct 1, 2026 •

Copy link
Copy Markdown
Contributor

A session save wrote back the whole session as the request loaded it. When requests on one session overlap, the one that finishes later puts its stale copy back over the other's changes, and re-creates the session if the other request destroyed it. regenerate() moved the same stale copy to the new id.

Change. A save re-reads the stored session and applies only the keys this request set, changed or removed since it loaded; after clear() it replaces the session. If the stored session is gone, it stays gone: nothing is written and no cookie is sent. If the store cannot be read, nothing is written and the change is kept for a retry. regenerate() re-reads the same way before moving the session, and returns None instead of minting an id for a session that has ended.

Test. tests/test_session_concurrent_requests.py: two sessions on one store stand in for overlapping requests (27 cases, including the documented login flow and an SSO callback). 16 fail on v3 without the change. Each part of the change, reverted on its own, turns a case red. The full suite shows no new failures against v3 on the same machine. Also run end to end over HTTP on the built-in server, with the file and the database (SQLite) session backends.

Parity. The same change in all four frameworks: #189, tina4stack/tina4-php#260, tina4stack/tina4-ruby#94, tina4stack/tina4-nodejs#108.

…ting it

A save wrote back the whole session as this request loaded it. When
requests on one session overlapped, the one that finished later put its
stale copy back over the other's changes, and re-created the session if
the other request had destroyed it. regenerate() moved the same stale
copy to the new id.

A save now re-reads the stored session and applies only the keys this
request set, changed or removed since it loaded; after clear() it
replaces the session. If the stored session is gone, it stays gone:
nothing is written and no cookie is sent. If the store cannot be read,
nothing is written and the change is kept for a retry. regenerate()
re-reads the same way before moving the session, and mints no id for a
session that has ended.

Signed-off-by: Michael <[email protected]>
@MichaelC8E

Copy link
Copy Markdown
Contributor Author

I have read the Tina4 Contributor Licence Agreement and I agree to it.

@MichaelC8E MichaelC8E closed this Oct 1, 2026
@MichaelC8E MichaelC8E reopened this Oct 1, 2026
MichaelC8E and others added 2 commits October 1, 2026 16:44
…session merge

The merge on save adds complexity to this one file. Only its entry in the baseline changes.

Signed-off-by: Michael <[email protected]>
…rrent-request tests

save() now re-writes a stored session even when this request changed nothing,
re-stamping the backend deadline to now + TINA4_SESSION_TTL so a session expires
after inactivity, not a fixed span after its last change. The re-write is the
same re-read, merged record the concurrent-save contract computes, so sliding
stays concurrent-safe and never re-creates an ended record. All existing guards
keep their precedence: no session id writes nothing; an ended (gone) record is
forgotten, not re-created; a failed backend read keeps the dirty flag for a
retry; a cleared session replaces rather than merges; TINA4_SESSION_TTL=0 still
means never-expires. PHP is the reference (ADR-0087).

tests/test_session_concurrent_requests.py: the two monkeypatch tests are
replaced with real-dependency tests (no mock/monkeypatch anywhere in the file):
- a read-only request now MOVES THE EXPIRY FORWARD - a real FileSessionHandler
  on a real tmp dir, the stored _expires read off disk before/after, asserting
  it advanced while the stored data is unchanged.
- the unreachable-store test points a REAL RedisSessionHandler at a closed TCP
  port (bind to get a free port, close it) so the backend genuinely refuses;
  save() returns False and the dirty change is retained for a later retry.

Signed-off-by: Andre van Zuydam <[email protected]>
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Co-Authored-By: Tina4 <[email protected]>
@tina4stack
tina4stack merged commit ff02c74 into tina4stack:v3 Oct 1, 2026
11 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants