GitLab

Even though it should not raise, a stale lock would be a serious issue.

The intent is to be able to tell that an independently-defined group of
activity nodes may execute given activity, and no other node.
This allows more flexible parallelism control than serialization_tag.

The fixes a missing rename in
commit cee3e728.

The goal is to make better use of the ZODB Storage cache. It is common to do
processing on a data set in several sequential transactions: in such case, by
continuing execution of these messages on the same node, data is loaded from
ZODB only once. Without this, and if there are many other messages to process,
processing always continue on a random node, causing much more load from ZODB.

To prevent nodes from having too much work to do, or too little compared to
other nodes, this new parameter is only a hint for CMFActivity. It remains
possible for a node to execute a message that was intended for another node.

Before this commit, a processing node selects the first message(s) according to
the following ordering:

  priority, date

and now:

  priority, node_preference, date

where node_preference is:

  -1 -> same node
   0 -> no preferred node
   1 -> another node

The implementation is tricky for 2 reasons:
- MariaDB can't order this way in a single simple query, so we have 1
  subquery for each case, potentially getting 3 times the wanted maximum of
  messages, then order/filter on the resulting union.
- MariaDB also can't filter efficiently messages for other nodes, so the 3rd
  subquery returns messages for any node, potentially duplicating results from
  the first 2 subqueries. This works because they'll be ordered last.
  Unfortunately, this requires extra indices.

In any case, message reservation must be very efficient, or MariaDB deadlocks
quickly happen, and locking an activity table during reservation reduces
parallelism too much.

In addition to better cache efficiency, this new feature can be used as a
workaround for a bug affecting serialiation_tag, causing IntegrityError when
reindexing many new objects. If you have 2 recursive reindexations for both a
document and one of its lines, and if you have so many messages than grouping
is split between these 2 messages, then you end up with 2 nodes indexing the
same line in parallel: for some tables, the pattern DELETE+INSERT conflicts
since InnoDB does not take any lock when deleting a non-existent row.

If you have many activities creating such documents, you can combine with
grouping and appropriate priority to make sure that such pair of messages won't
be executed on different nodes, except maybe at the end (when there's no
document to create anymore; then activity reexecution may be enough).
For example:

  from Products.CMFActivity.ActivityTool import getCurrentNode
  portal.setPlacelessDefaultReindexParameters(
    activate_kw={'node': 'same', 'priority': priority},
    group_id=getCurrentNode())

where `priority` is the same as the activity containing the above code, which
can also use grouping without increasing the probability of IntegrityError.

The original goal was to improve performance by removing the
`processing_node_processing` index and the queries that modified
these 2 useless columns.

As shown in the following example, on a big catalog table,
MariaDB is able to use several indices at the same time
('...' are obfuscated unique values):

  > analyze select SQL_NO_CACHE uid, relative_url from catalog where reference='...' OR relative_url='...';
  +------+-------------+---------+-------------+------------------------+------------------------+---------+------+------+--------+----------+------------+--------------------------------------------------+
  | id   | select_type | table   | type        | possible_keys          | key                    | key_len | ref  | rows | r_rows | filtered | r_filtered | Extra                                            |
  +------+-------------+---------+-------------+------------------------+------------------------+---------+------+------+--------+----------+------------+--------------------------------------------------+
  |    1 | SIMPLE      | catalog | index_merge | Reference,relative_url | Reference,relative_url | 768,767 | NULL |    2 |   2.00 |   100.00 |     100.00 | Using union(Reference,relative_url); Using where |
  +------+-------------+---------+-------------+------------------------+------------------------+---------+------+------+--------+----------+------------+--------------------------------------------------+
  1 row in set (0.00 sec)

So mixing different dependency types with OR should be fine
(no need to split into more subqueries and join with UNION).

Reduces code duplication.

portal_activities contains documents which need to be (and are) indexed.
For consistency, the tool itself should be indexable.

- getCurrentNode method can be directly imported from CMFActivity
- corresponding unittests

/reviewed-on nexedi/erp5!647

This commit:

- Adds a new Activity called "SQLJoblib"
- Adds a Backend to be used by joblib
- Uses OOBTree to store results instead of ConflictFreeLog
- Adds a getResultDict API to fetch resut Dict

It uses the original work from rafael@nexedi.com and loic.esteve@inria.fr

otherwise Localizer will not work in grouped activity.

This makes a difference for SQLDict as not all messages are accepted for
insertion.

For the new GroupedMessage class in commit
da234001, I renamed 'obj' to 'object'
at the last minute and I missed 2 occurrences.

The recent API change was not enough.
A grouping method may need more information: in particular, the dummy grouping
method must be fixed to change user.

When grouped messages fail, ActivityTool must distinguish 3 groups,
in order to reexecute them separately, as follows:
- first, those that succeeded
- then, those that were skipped
- at last, failed ones

Grouping methods are updated to handle partial failures, and stop doing
anything when something goes wrong.

Without this, we would have the following pathological cases.

1. Let's suppose first that skipped messages are marked as succeeded.

The problem is that each skipped message that will fail causes the reexecution
of those that didn't fail.

Exemple: A:ok B:ok C:err D:err E:err F:err
  1: A:ok, B:ok, C:err, D:skipped, E:skipped, F:skipped
  2: A:ok, B:ok, D:err, E:skipped, F:skipped
  3: A:ok, B:ok, E:err, F:skipped
  4: A:ok, B:ok, F:err
  5: A:ok, B:ok -> commit

And worst, the first failed (C) may be processable again before 5, entering
a failing loop if it is executed again in the same group as A & B.

2. Another implementation is to mar...

For even more refactoring between SQLDict & SQLQueue,
which now uses SQL tables with the same schema.

The action to recreate activity tables while preserving existing messages
was unsafe for 2 reasons:
- if any error happened, messages could be lost
- it relied on Message.reactivate

Which this patch, any instance created after commit d881edd1 (Aug 2010) will
upgrade successfully. For older instances, make sure you have no activity left.

For cases where 'ALTER TABLE' would not work, a better way to implement repair
functionality would be:
- one action to backup all messages in ZODB
- and another to restore them
And maybe a security so that during the backup-clear-restore sequence,
activities can't be created nor processed.

If any column is added in the future, it would still be possible to write code
that fills them by inspecting messages.

This reverts commit 6a8987fb.

Assumes nodes will not be named after other node's network address.