qpid-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Pavel Moravec (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (QPID-5880) [C++ broker] Make memory usage consistent after broker restart
Date Mon, 07 Jul 2014 14:14:38 GMT

     [ https://issues.apache.org/jira/browse/QPID-5880?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Pavel Moravec updated QPID-5880:
--------------------------------

    Attachment: QPID-5880.patch

Simple idea:
- in Queue::enqueue, set PersistableMessage::persistencyID "manually" to some unique number
that is identical to all message instances that has common SharedState - e.g. to the pointer
to SharedState
- during journal recovery, if we recover a message with already seen persistencyID, use the
previous seen instead with its SharedState and PersistableMessage bits

Known limitation:
- message annotations added to some queue (e.g. due to queue sequencing enabled) will be either
over-written or shared to all other queues during recovery

The patch contains a new QueueSettings option to enable (bu default disabled) this feature
on per queue basis. This somehow limits the limitation above.

Isn't storing pointer to SharedState to the disk (via persistencyID) some sort of security
breach? (I dont think so, but worth to ask)

Can't manual setup of persistencyID break something in store? (AFAIK no as uniqueness of the
ID is assured: 1) a new / different message with the same persistencyID can appear only after
the previous instance is gone from memory, and 2) only queues with the option enabled are
checked for message coupling)

Will it work well in cluster? No, it won't. As when primary broker gets 1 message to an exchange
that distributes it to 100 queues, th broker updates backup brokers via 100 individual "enqueue
1 message to queue q[1-100]" events. So backup brokers consume more memory than primary -
the same amount like primary does not share SharedState at all.

Not sure if the patch is reasonable due to the cluster/HA point.

> [C++ broker] Make memory usage consistent after broker restart
> --------------------------------------------------------------
>
>                 Key: QPID-5880
>                 URL: https://issues.apache.org/jira/browse/QPID-5880
>             Project: Qpid
>          Issue Type: Improvement
>          Components: C++ Broker
>    Affects Versions: 0.28
>            Reporter: Pavel Moravec
>            Assignee: Pavel Moravec
>            Priority: Minor
>         Attachments: QPID-5880.patch
>
>
> In scenario "send messages via (fanout) exchange into multiple (tens, hundreds) queues",
qpid broker keeps SharedState just once per the received message, to save RAM requirements.
> Assuming the messages are durable and stored to the disk, broker restart and journal
recovery creates one SharedState for each message copy in each queue. This has negative impact
to memory consumed by the broker, and it could potentially prevent broker successfull startup
due to RAM exhausted - while the broker was running fine prior the restart attempt.
> It is required to have some "unifying" mechanism in journal recovery that creates just
one SharedState per identical copies of a message spread over more queues.
> Ideally, this could be resolved by durable topics but that would require much bigger
change.
> My proposal (to be discussed in reviewboard) is rather a fast fix supplying the original
demand.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@qpid.apache.org
For additional commands, e-mail: dev-help@qpid.apache.org


Mime
View raw message