[
https://issues.apache.org/jira/browse/DERBY-4437?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mike Matrigali updated DERBY-4437:
----------------------------------
Here are my thoughts. The work so far looks great and I would be fine seeing
it checked in as is and backported. I think it is
very reasonable to log a JIRA for the tunable aspect and hope someone is
interested in that work. I actually think that is better as I would
rather see the tunable aspect go into only trunk and a new release rather than
backported as a bug fix - since as you point out it is experimental.
It would be great if we could have a discussion there on possible ideas for an
algorithm there. I actually think the project could be done by
a newcomer as the coding would be very localized once you check in, the hard
part is just coding the tuning stuff in one place. Might have
to do some new work to have interesting inputs to the algorithm.
It may also be worth logging a separate improvement JIRA if one does not exist
to solve the lost range on JIRA crash. This is not a simple problem, and may
not be worth effort but might as well create a place holder. It would somehow
probably need new log records and specialized recovery of those during crash
recovery. It is complicated in that it is logical work that needs to be done
above store but currently only store work is done during crash recovery.
Another
option would be some brute force work to actually scan the tables on recovery
and find the "highest" and reset the range before an application came in. If
there are indexes on these things this becomes much faster. Some of this work
is similar to the problem we have with post commit space reclamation
work that is also lost when a crash happens.
The following would be my votes on the changes you mentioned you are willing to
do, but would not argue strongly with opposing views:
o if a knob is added so a user could backtrack if they did not like losing
ranges, then I would bump the default to something like 200, given
knut's experiments. I would backport this change to 10.8, again only if
knob was included. Given how fast processors are nowadays adding
a commit I/O for every 5 inserts seems a high price, when the system is
doing 900 inserts a second.
o If no knob I think I would leave default of 5 for backport to 10.8 and bump
default to 200 in trunk. We could give notice in 10.8 release that upcoming
10.9 would bump this default.
o Using the properties seems like complicated syntax to me. I assume using
"alter table" is not possible as there is no standard in this area. What would
people
think about using a system procedure instead of a property. That way the
call could simply take the standard application and table name arguments, and
would require the usual alter table database permissions to set. Underneath
the procedure call it should just call an internal alter table call.
> Concurrent inserts into table with identity column perform poorly
> -----------------------------------------------------------------
>
> Key: DERBY-4437
> URL: https://issues.apache.org/jira/browse/DERBY-4437
> Project: Derby
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 10.5.3.0
> Reporter: Knut Anders Hatlen
> Assignee: Rick Hillegas
> Attachments: D4437PerfTest.java, D4437PerfTest2.java,
> Experiments_4437.html, derby-4437-01-aj-allTestsPass.diff,
> derby-4437-02-ac-alterTable-bulkImport-deferredInsert.diff,
> derby-4437-03-aa-upgradeTest.diff,
> derby-4437-04-aa-reclaimUnusedValuesOnShutdown.diff,
> derby-4437-05-aa-pluggablePreallocation.diff,
> derby-4437-06-aa-selfTuning.diff, insertperf.png, insertperf2.png,
> prealloc.png
>
>
> I have a multi-threaded application which is very insert-intensive. I've
> noticed that it sometimes can come into a state where it slows down
> considerably and basically becomes single-threaded. This is especially
> harmful on modern multi-core machines since most of the available resources
> are left idle.
> The problematic tables contain identity columns, and here's my understanding
> of what happens:
> 1) Identity columns are generated from a counter that's stored in a row in
> SYS.SYSCOLUMNS. During normal operation, the counter is maintained in a
> nested transaction within the transaction that performs the insert. This
> allows the nested transaction to commit the changes to SYS.SYSCOLUMN
> separately from the main transaction, and the exclusive lock that it needs to
> obtain on the row holding the counter, can be releases after a relatively
> short time. Concurrent transactions can therefore insert into the same table
> at the same time, without needing to wait for the others to commit or abort.
> 2) However, if the nested transaction cannot lock the row in SYS.SYSCOLUMNS
> immediately, it will give up and retry the operation in the main transaction.
> This prevents self-deadlocks in the case where the main transaction already
> owns a lock on SYS.SYSCOLUMNS. Unfortunately, this also increases the time
> the row is locked, since the exclusive lock cannot be released until the main
> transaction commits. So as soon as there is one lock collision, the waiting
> transaction changes to a locking mode that increases the chances of others
> having to wait, which seems to result in all insert threads having to obtain
> the SYSCOLUMNS locks in the main transaction. The end result is that only one
> of the insert threads can execute at any given time as long as the
> application is in this state.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
