yahoo-eng-team team mailing list archive

[Dolph Mathews <1623117@xxxxxxxxxxxxxxxxxx>]

Public bug reported:

There are three scenarios during a rolling upgrade process where we
could prevent operators from doing the "wrong thing" (doing things out
of order):

1) Operators running code from the next release before `keystone-manage
db_sync --expand` has been run: If you run the next release before
--expand is run, then you'll surely end up with fatal query errors as
columns and tables won't exist that the app thinks should exist.

2) (the scary one) Operators running code from the next release before
`keystone-manage db_sync --migrate` has been run: If you run the next
release before --migrate is run, then any number of different types of
failures are possible due to unpopulated columns & tables, including a
risk of data loss as the new release tries to update records that it
perceives to be unpopulated, which might propagate to the legacy schema
during UPDATE operations, for example.

3) Operators running code from the previous release after `keystone-
manage db_sync --contract` has been run: As in case (1), this may result
in fatal query errors, but also presents a risk of introducing data
inconsistency, as the legacy schema might not have a "full
understanding" of the new schema, as would be the case with additive
schema changes. The legacy application would no longer have triggers to
rely on, so consequences would mostly be dependent on the default values
of columns, constraints, etc.

The second case worries me, as it's the most likely scenario where
operators might not realize what's going on until it's too late.

To prevent all of these scenarios, I think the application should check
at startup to ensure that the expand and data migration repositories
both match a minimum value (specifically, the most recent migration in
the application's respective repositories).

Doing the same sort of check at startup for the contract repo would be
more difficult, as it'd be entirely dependent on when you last upgraded
(whether it be last stable/* release or master at any point), so I'd
like to leave that out of scope here.

** Affects: keystone
     Importance: Medium
         Status: New


** Tags: upgrades

** Summary changed:

- Prevent keystone from serving requests when schema is not up to date
+ Prevent keystone from serving requests when schema or data migrations are not up to date

-- 
You received this bug notification because you are a member of Yahoo!
Engineering Team, which is subscribed to OpenStack Identity (keystone).
https://bugs.launchpad.net/bugs/1623117

Title:
  Prevent keystone from serving requests when schema or data migrations
  are not up to date

Status in OpenStack Identity (keystone):
  New

Bug description:
  There are three scenarios during a rolling upgrade process where we
  could prevent operators from doing the "wrong thing" (doing things out
  of order):

  1) Operators running code from the next release before `keystone-
  manage db_sync --expand` has been run: If you run the next release
  before --expand is run, then you'll surely end up with fatal query
  errors as columns and tables won't exist that the app thinks should
  exist.

  2) (the scary one) Operators running code from the next release before
  `keystone-manage db_sync --migrate` has been run: If you run the next
  release before --migrate is run, then any number of different types of
  failures are possible due to unpopulated columns & tables, including a
  risk of data loss as the new release tries to update records that it
  perceives to be unpopulated, which might propagate to the legacy
  schema during UPDATE operations, for example.

  3) Operators running code from the previous release after `keystone-
  manage db_sync --contract` has been run: As in case (1), this may
  result in fatal query errors, but also presents a risk of introducing
  data inconsistency, as the legacy schema might not have a "full
  understanding" of the new schema, as would be the case with additive
  schema changes. The legacy application would no longer have triggers
  to rely on, so consequences would mostly be dependent on the default
  values of columns, constraints, etc.

  The second case worries me, as it's the most likely scenario where
  operators might not realize what's going on until it's too late.

  To prevent all of these scenarios, I think the application should
  check at startup to ensure that the expand and data migration
  repositories both match a minimum value (specifically, the most recent
  migration in the application's respective repositories).

  Doing the same sort of check at startup for the contract repo would be
  more difficult, as it'd be entirely dependent on when you last
  upgraded (whether it be last stable/* release or master at any point),
  so I'd like to leave that out of scope here.

To manage notifications about this bug go to:
https://bugs.launchpad.net/keystone/+bug/1623117/+subscriptions