Moved much of the writing that happens when posting a notice to a new
queuehandler, distribqueuehandler. This updates tags, groups, replies
and inboxes at queue time (or at Web time, if queues are disabled).
To make this work well, I had to break up the monolithic
Notice::blowCaches() and make cache blowing happen closer to where
data is updated.
Squashed commit of the following:
commit 5257626c62750ac4ac1db0ce2b71410c5711cfa3
Author: Evan Prodromou <evan@status.net>
Date: Mon Jan 25 14:56:41 2010 -0500
slightly better handling of blowing tag memory cache
commit 8a22a3cdf6ec28685da129a0313e7b2a0837c9ef
Author: Evan Prodromou <evan@status.net>
Date: Mon Jan 25 01:42:56 2010 -0500
change 'distribute' to 'distrib' so not too long for dbqueue
commit 7a063315b0f7fad27cb6fbd2bdd74e253af83e4f
Author: Evan Prodromou <evan@status.net>
Date: Mon Jan 25 01:39:15 2010 -0500
change handle_notice() to handle() in distributqueuehandler
commit 1a39ccd28b9994137d7bfd21bb4f230546938e77
Author: Evan Prodromou <evan@status.net>
Date: Mon Jan 25 16:05:25 2010 -0500
error with queuemanager
commit e6b3bb93f305cfd2de71a6340b8aa6fb890049b7
Author: Evan Prodromou <evan@status.net>
Date: Mon Jan 25 01:11:34 2010 -0500
Blow memcache at different point rather than one big function for Notice class
commit 94d557cdc016187d1d0647ae1794cd94d6fb8ac8
Author: Evan Prodromou <evan@status.net>
Date: Mon Jan 25 00:48:44 2010 -0500
Blow memcache at different point rather than one big function for Notice class
commit 1c781dd08c88a35dafc5c01230b4872fd6b95182
Author: Evan Prodromou <evan@status.net>
Date: Wed Jan 20 08:54:18 2010 -0500
move broadcasting and distributing to new queuehandler
commit da3e46d26b84e4f028f34a13fd2ee373e4c1b954
Author: Evan Prodromou <evan@status.net>
Date: Wed Jan 20 08:53:12 2010 -0500
Move distribution of notices to new distribute queue handler
Previously, messages once delivered would just get stuck in the queue seemingly forever if they never got ACKed.
Note this could lead to partial duplication, for instance if the OMB or Twitter queue handlers die after 1/2 of the outgoing sends.
Recommendations:
* catch exceptions more aggressively within queue handlers (so only PHP fatal errors are likely to kill in the middle)
* for processing that involves sending to multiple clients, consider a second queue similar to the XMPP output, eg for OMB:
- first queue gets delivery list and builds message data, enqueueing it for each target address
- second queue can handle each individual outgoing message (and attempt redelivery etc separately)
This would also protect better against a recurring error preventing delivery in the second part, and could spread out any slow sends over multiple threads.
Queue handlers for XMPP individual & firehose output now send their XML stanzas
to another output queue instead of connecting directly to the chat server. This
lets us have as many general processing threads as we need, while all actual
XMPP input and output go through a single daemon with a single connection open.
This avoids problems with multiple connected resources:
* multiple windows shown in some chat clients (psi, gajim, kopete)
* extra load on server
* incoming message delivery forwarding issues
Database changes:
* queue_item drops 'notice_id' in favor of a 'frame' blob.
This is based on Craig Andrews' work branch to generalize queues to take any
object, but conservatively leaving out the serialization for now.
Table updater (preserves any existing queued items) in db/rc3to09.sql
Code changes to watch out for:
* Queue handlers should now define a handle() method instead of handle_notice()
* QueueDaemon and XmppDaemon now share common i/o (IoMaster) and respawning
thread management (RespawningDaemon) infrastructure.
* The polling XmppConfirmManager has been dropped, as the message is queued
directly when saving IM settings.
* Enable $config['queue']['debug_memory'] to output current memory usage at
each run through the event loop to watch for memory leaks
To do:
* Adapt XMPP i/o to component connection mode for multi-site support.
* XMPP input can also be broken out to a queue, which would allow the actual
notice save etc to be handled by general queue threads.
* Make sure there are no problems with simply pushing serialized Notice objects
to queues.
* Find a way to improve interactive performance of the database-backed queue
handler; polling is pretty painful to XMPP.
* Possibly redo the way QueueHandlers are injected into a QueueManager. The
grouping used to split out the XMPP output queue is a bit awkward.
Conflicts:
scripts/xmppdaemon.php
Previously, messages once delivered would just get stuck in the queue seemingly forever if they never got ACKed.
Note this could lead to partial duplication, for instance if the OMB or Twitter queue handlers die after 1/2 of the outgoing sends.
Recommendations:
* catch exceptions more aggressively within queue handlers (so only PHP fatal errors are likely to kill in the middle)
* for processing that involves sending to multiple clients, consider a second queue similar to the XMPP output, eg for OMB:
- first queue gets delivery list and builds message data, enqueueing it for each target address
- second queue can handle each individual outgoing message (and attempt redelivery etc separately)
This would also protect better against a recurring error preventing delivery in the second part, and could spread out any slow sends over multiple threads.
Queue handlers for XMPP individual & firehose output now send their XML stanzas
to another output queue instead of connecting directly to the chat server. This
lets us have as many general processing threads as we need, while all actual
XMPP input and output go through a single daemon with a single connection open.
This avoids problems with multiple connected resources:
* multiple windows shown in some chat clients (psi, gajim, kopete)
* extra load on server
* incoming message delivery forwarding issues
Database changes:
* queue_item drops 'notice_id' in favor of a 'frame' blob.
This is based on Craig Andrews' work branch to generalize queues to take any
object, but conservatively leaving out the serialization for now.
Table updater (preserves any existing queued items) in db/rc3to09.sql
Code changes to watch out for:
* Queue handlers should now define a handle() method instead of handle_notice()
* QueueDaemon and XmppDaemon now share common i/o (IoMaster) and respawning
thread management (RespawningDaemon) infrastructure.
* The polling XmppConfirmManager has been dropped, as the message is queued
directly when saving IM settings.
* Enable $config['queue']['debug_memory'] to output current memory usage at
each run through the event loop to watch for memory leaks
To do:
* Adapt XMPP i/o to component connection mode for multi-site support.
* XMPP input can also be broken out to a queue, which would allow the actual
notice save etc to be handled by general queue threads.
* Make sure there are no problems with simply pushing serialized Notice objects
to queues.
* Find a way to improve interactive performance of the database-backed queue
handler; polling is pretty painful to XMPP.
* Possibly redo the way QueueHandlers are injected into a QueueManager. The
grouping used to split out the XMPP output queue is a bit awkward.
Queue handlers for XMPP individual & firehose output now send their XML stanzas
to another output queue instead of connecting directly to the chat server. This
lets us have as many general processing threads as we need, while all actual
XMPP input and output go through a single daemon with a single connection open.
This avoids problems with multiple connected resources:
* multiple windows shown in some chat clients (psi, gajim, kopete)
* extra load on server
* incoming message delivery forwarding issues
Database changes:
* queue_item drops 'notice_id' in favor of a 'frame' blob.
This is based on Craig Andrews' work branch to generalize queues to take any
object, but conservatively leaving out the serialization for now.
Table updater (preserves any existing queued items) in db/rc3to09.sql
Code changes to watch out for:
* Queue handlers should now define a handle() method instead of handle_notice()
* QueueDaemon and XmppDaemon now share common i/o (IoMaster) and respawning
thread management (RespawningDaemon) infrastructure.
* The polling XmppConfirmManager has been dropped, as the message is queued
directly when saving IM settings.
* Enable $config['queue']['debug_memory'] to output current memory usage at
each run through the event loop to watch for memory leaks
To do:
* Adapt XMPP i/o to component connection mode for multi-site support.
* XMPP input can also be broken out to a queue, which would allow the actual
notice save etc to be handled by general queue threads.
* Make sure there are no problems with simply pushing serialized Notice objects
to queues.
* Find a way to improve interactive performance of the database-backed queue
handler; polling is pretty painful to XMPP.
* Possibly redo the way QueueHandlers are injected into a QueueManager. The
grouping used to split out the XMPP output queue is a bit awkward.
- NOTICE_INBOX_SOURCE_* constants moved to common.php since Notice_inbox.php not always loaded
- fixed typo in User::staticGet() call which caused user #1 to receive messages once for each subscriber instead of for him/herself
- 'continue' -> 'continue 2' inside switch() statement to fix loop escape (PHP considers switch() a looping construct for break & continue)
Key changes:
* Initialization code moved from common.php to StatusNet class;
can now switch configurations during runtime.
* As a consequence, configuration files must now be idempotent...
Be careful with constant, function or class definitions.
* Control structure for daemons/QueueManager/QueueHandler has been refactored;
the run loop is now managed by IoMaster run via scripts/queuedaemon.php
IoManager subclasses are woken to handle socket input or polling, and may
cover multiple sites.
* Plugins can implement notice queue handlers more easily by registering a
QueueHandler class; no more need to add a daemon.
The new QueueDaemon runs from scripts/queuedaemon.php:
* This replaces most of the old *handler.php scripts; they've been refactored
to the bare handler classes.
* Spawns multiple child processes to spread load; defaults to CPU count on
Linux and Mac OS X systems, or override with --threads=N
* When multithreaded, child processes are automatically respawned on failure.
* Threads gracefully shut down and restart when passing a soft memory limit
(defaults to 90% of memory_limit), limiting damage from memory leaks.
* Support for UDP-based monitoring: http://www.gitorious.org/snqmon
Rough control flow diagram:
QueueDaemon -> IoMaster -> IoManager
QueueManager [listen or poll] -> QueueHandler
XmppManager [ping & keepalive]
XmppConfirmManager [poll updates]
Todo:
* Respawning features not currently available running single-threaded.
* When running single-site, configuration changes aren't picked up.
* New sites or config changes affecting queue subscriptions are not yet
handled without a daemon restart.
* SNMP monitoring output to integrate with general tools (nagios, ganglia)
* Convert XMPP confirmation message sends to use stomp queue instead of polling
* Convert xmppdaemon.php to IoManager?
* Convert Twitter status, friends import polling daemons to IoManager
* Clean up some error reporting and failure modes
* May need to adjust queue priorities for best perf in backlog/flood cases
Detailed code history available in my daemon-work branch:
http://www.gitorious.org/~brion/statusnet/brion-fixes/commits/daemon-work
* Mostly punctuation updates so that the same message is used consistently in all of StatusNet.
* Some cases of "Title Case" removed, because that does not appear to be used consistently.
This reverts commit 5d9a2eb17e.
These are commands that are/were implemented by Twitter, and we don't
(yet) implemented. People will be looking for that information.
* We now cache negative lookups; clear them in Memcached_DataObject->insert()
* Mark file.url as a unique key in statusnet.ini so its negative lookups are cleared properly (first save of a notice with a new URL was failing due to double-insert)
* Now using serialization for default in-process cache instead of just saving objects; avoids potential corruption if you save an object to cache, change the original object, then fetch the same key from cache again
Consolidated several separate implementations of the same weighting algorithm into common_sql_weight() and fixed some bugs...
For MySQL, now using timestampdiff() instead of subtraction for the comparison, so we get sane results when the year doesn't match, and utc_timestamp() rather than now() so we don't get negative ages for recent items with local server timezone.
Unknown whether the same problems affect PostgreSQL, but note that it lacks the timestampdiff() SQL function.
Consolidated several separate implementations of the same weighting algorithm into common_sql_weight() and fixed some bugs...
For MySQL, now using timestampdiff() instead of subtraction for the comparison, so we get sane results when the year doesn't match, and utc_timestamp() rather than now() so we don't get negative ages for recent items with local server timezone.
Unknown whether the same problems affect PostgreSQL, but note that it lacks the timestampdiff() SQL function.
Allows storage of larger objects (over 1mb in size uncompressed), such as huge LDAP schemas.
Should also improve cache efficiency (allows more stuff to be stored in same memory) and reduce network latency (less data transfer)