Open Stack

On 06/09/2014 12:50 PM, Devananda van der Veen wrote:
> There may be some problems with MySQL when testing parallel writes in
> different non-committing transactions, even in READ COMMITTED mode,
> due to InnoDB locking, if the queries use non-unique secondary indexes
> for UPDATE or SELECT..FOR UPDATE queries. This is done by the
> "with_lockmode('update')" SQLAlchemy phrase, and is used in ~10 places
> in Nova. So I would not recommend this approach, even though, in
> principle, I agree it would be a much more efficient way of testing
> database reads/writes.
>
> More details here:
> http://dev.mysql.com/doc/refman/5.5/en/innodb-locks-set.html and
> http://dev.mysql.com/doc/refman/5.5/en/innodb-record-level-locks.html

Hi Deva,

MySQL/InnoDB's default isolation mode is REPEATABLE_READ, not 
READ_COMMITTED... are you saying that somewhere in the Ironic codebase 
we are setting the isolation mode manually to READ_COMMITTED for some 
reason?

Best,
-jay

> On Sun, Jun 8, 2014 at 8:46 AM, Roman Podoliaka <rpodolyaka at mirantis.com> wrote:
>> Hi Mike,
>>
>>>>> However, when testing an application that uses a fixed set of tables, as should be the case for the majority if not all Openstack apps, there’s no reason that these tables need to be recreated for every test.
>>
>> This is a very good point. I tried to use the recipe from SQLAlchemy
>> docs to run Nova DB API tests (yeah, I know, this might sound
>> confusing, but these are actually methods that access the database in
>> Nova) on production backends (MySQL and PostgreSQL). The abandoned
>> patch is here [1]. Julia Varlamova has been working on rebasing this
>> on master and should upload a new patch set soon.
>>
>> Overall, the approach with executing a test within a transaction and
>> then emitting ROLLBACK worked quite well. The only problem I ran into
>> were tests doing ROLLBACK on purpose. But you've updated the recipe
>> since then and this can probably be solved by using of save points. I
>> used a separate DB per a test running process to prevent race
>> conditions, but we should definitely give READ COMMITTED approach a
>> try. If it works, that will awesome.
>>
>> With a few tweaks of PostgreSQL config I was able to run Nova DB API
>> tests in 13-15 seconds, while SQLite in memory took about 7s.
>>
>> Action items for me and Julia probably: [2] needs a spec with [1]
>> updated accordingly. Using of this 'test in a transaction' approach
>> seems to be a way to go for running all db related tests except the
>> ones using DDL statements (as any DDL statement commits the current
>> transaction implicitly on MySQL and SQLite AFAIK).
>>
>> Thanks,
>> Roman
>>
>> [1] https://review.openstack.org/#/c/33236/
>> [2] https://blueprints.launchpad.net/nova/+spec/db-api-tests-on-all-backends
>>
>> On Sat, Jun 7, 2014 at 10:27 PM, Mike Bayer <mbayer at redhat.com> wrote:
>>>
>>> On Jun 6, 2014, at 8:12 PM, Devananda van der Veen <devananda.vdv at gmail.com>
>>> wrote:
>>>
>>> I think some things are broken in the oslo-incubator db migration code.
>>>
>>> Ironic moved to this when Juno opened and things seemed fine, until recently
>>> when Lucas tried to add a DB migration and noticed that it didn't run... So
>>> I looked into it a bit today. Below are my findings.
>>>
>>> Firstly, I filed this bug and proposed a fix, because I think that tests
>>> that don't run any code should not report that they passed -- they should
>>> report that they were skipped.
>>>    https://bugs.launchpad.net/oslo/+bug/1327397
>>>    "No notice given when db migrations are not run due to missing engine"
>>>
>>> Then, I edited the test_migrations.conf file appropriately for my local
>>> mysql service, ran the tests again, and verified that migration tests ran --
>>> and they passed. Great!
>>>
>>> Now, a little background... Ironic's TestMigrations class inherits from
>>> oslo's BaseMigrationTestCase, then "opportunistically" checks each back-end,
>>> if it's available. This opportunistic checking was inherited from Nova so
>>> that tests could pass on developer workstations where not all backends are
>>> present (eg, I have mysql installed, but not postgres), and still
>>> transparently run on all backends in the gate. I couldn't find such
>>> opportunistic testing in the oslo db migration test code, unfortunately -
>>> but maybe it's well hidden.
>>>
>>> Anyhow. When I stopped the local mysql service (leaving the configuration
>>> unchanged), I expected the tests to be skipped, but instead I got two
>>> surprise failures:
>>> - test_mysql_opportunistically() failed because setUp() raises an exception
>>> before the test code could call calling _have_mysql()
>>> - test_mysql_connect_fail() actually failed! Again, because setUp() raises
>>> an exception before running the test itself
>>>
>>> Unfortunately, there's one more problem... when I run the tests in parallel,
>>> they fail randomly because sometimes two test threads run different
>>> migration tests, and the setUp() for one thread (remember, it calls
>>> _reset_databases) blows up the other test.
>>>
>>> Out of 10 runs, it failed three times, each with different errors:
>>>    NoSuchTableError: `chassis`
>>>    ERROR 1007 (HY000) at line 1: Can't create database 'test_migrations';
>>> database exists
>>>    ProgrammingError: (ProgrammingError) (1146, "Table
>>> 'test_migrations.alembic_version' doesn't exist")
>>>
>>> As far as I can tell, this is all coming from:
>>>
>>> https://github.com/openstack/oslo-incubator/blob/master/openstack/common/db/sqlalchemy/test_migrations.py#L86;L111
>>>
>>>
>>> Hello -
>>>
>>> Just an introduction, I’m Mike Bayer, the creator of SQLAlchemy and Alembic
>>> migrations.     I’ve just joined on as a full time Openstack contributor,
>>> and trying to help improve processes such as these is my primary
>>> responsibility.
>>>
>>> I’ve had several conversations already about how migrations are run within
>>> test suites in various openstack projects.   I’m kind of surprised by this
>>> approach of dropping and recreating the whole database for individual tests.
>>> Running tests in parallel is obviously made very difficult by this style,
>>> but even beyond that, a lot of databases don’t respond well to lots of
>>> dropping/rebuilding of tables and/or databases in any case; while SQLite and
>>> MySQL are probably the most forgiving of this, a backend like Postgresql is
>>> going to lock tables from being dropped more aggressively, if any open
>>> transactions or result cursors against those tables remain, and on a backend
>>> like Oracle, the speed of schema operations starts to become prohibitively
>>> slow.   Dropping and creating tables is in general not a very speedy task on
>>> any backend, and on a test suite that runs many tests against a fixed
>>> schema, I don’t see why a full drop is necessary.
>>>
>>> If you look at SQLAlchemy’s own tests, they do in fact create tables on each
>>> test, or just as often for a specific suite of tests.  However, this is due
>>> to the fact that SQLAlchemy tests are testing SQLAlchemy itself, so the
>>> database schemas used for these tests are typically built explicitly for
>>> small groups or individual tests, and there are ultimately thousands of
>>> small “mini schemas” built up and torn down for these tests.   A lot of
>>> framework code is involved within the test suite to keep more picky
>>> databases like Postgresql and Oracle happy when building up and dropping
>>> tables so frequently.
>>>
>>> However, when testing an application that uses a fixed set of tables, as
>>> should be the case for the majority if not all Openstack apps, there’s no
>>> reason that these tables need to be recreated for every test.   Typically,
>>> the way I recommend is that the test suite includes a “per suite” activity
>>> which creates the test schema just once (with or without using CREATE
>>> DATABASE; I’m not a fan of tests running CREATE DATABASE as this is not a
>>> command so easily available in some environments).   The tests themselves
>>> then run within a transactional container, such that each test performs all
>>> of its work within a context that doesn’t actually commit any data to the
>>> database; a test that actually states “session.commit()” runs within a
>>> container that doesn’t actually emit the COMMIT, and if support is needed
>>> for tests that also emit “session.rollback()”, the container can be written
>>> to support this paradigm as well (I helped some Dropbox devs with such an
>>> approach recently).
>>>
>>> In this way, the database migrations are exercised, but only once at the
>>> beginning in order to build up the schema; the tests can then run with very
>>> low complexity/performance overhead as far as database-level setup/teardown,
>>> and parallel testing is also much easier.   When the test suite completes is
>>> when a drop of the entire set of tables can proceed.  Because tests are run
>>> within transactions, assuming READ COMMITTED isolation is established, the
>>> tests don’t even see the data being incurred by other tests running in
>>> parallel.
>>>
>>> It remains to be seen what aspects of Openstack I’m going to get involved
>>> with first, though this migration and testing issue seems to be a big one.
>>> I’d love to get comments from the community here as to how this process
>>> might be improved and if a rearrangement of the fixture system in the way I
>>> describe might be helpful and feasible.
>>>
>>> - mike
>>>
>>>
>>>
>>> _______________________________________________
>>> OpenStack-dev mailing list
>>> OpenStack-dev at lists.openstack.org
>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>>
>>
>> _______________________________________________
>> OpenStack-dev mailing list
>> OpenStack-dev at lists.openstack.org
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>