Comment 42 for bug 1844929

Revision history for this message
OpenStack Infra (hudson-openstack) wrote : Fix merged to nova (stable/stein)

Reviewed: https://review.opendev.org/720587
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=4de766006d9432a7ccbcf6a4d4232db472b2f0e1
Submitter: Zuul
Branch: stable/stein

commit 4de766006d9432a7ccbcf6a4d4232db472b2f0e1
Author: melanie witt <email address hidden>
Date: Fri Apr 3 21:22:27 2020 +0000

    Reset the cell cache for database access in Service

    We have had a gate bug for a long time where occasionally the scheduler
    service gets into a state where many requests fail in it with
    CellTimeout errors. Example:

      Timed out waiting for response from cell <cell uuid>

    Through the use of much DNM patch debug logging in oslo.db, it was
    revealed that service child processes (workers) were sometimes starting
    off with already locked internal oslo.db locks. This is a known issue
    in python [1] where if a parent process forks a child process while a
    lock is held, the child will inherit the held lock which can never be
    acquired.

    The python issue is not considered a bug and the recommended way to
    handle it is by making use of the os.register_at_fork() in the oslo.db
    to reinitialize its lock. The method is new in python 3.7, so as long
    as we still support python 3.6, we must handle the situation outside of
    oslo.db.

    We can do this by clearing the cell cache that holds oslo.db database
    transaction context manager objects during service start(). This way,
    we get fresh oslo.db locks that are in an unlocked state when a child
    process begins.

    We can also take this opportunity to resolve part of a TODO to clear
    the same cell cache during service reset() (SIGHUP) since it is another
    case where we intended to clear it. The rest of the TODO related to
    periodic clearing of the cache is removed after discussion on the
    review, as such clearing would be unsynchronized among multiple
    services and for periods of time each service might have a different
    view of cached cells than another.

    Closes-Bug: #1844929

    [1] https://bugs.python.org/issue6721

    Change-Id: Id233f673a57461cc312e304873a41442d732c051
    (cherry picked from commit 941559042f609ee43ff3160c0f0d0c45187be17f)
    (cherry picked from commit 88205a4e911268dae7120a6a43ff9042d1534251)