[drupal-devel] [feature] Alternative caching for high traffic websites

Dries drupal-devel at drupal.org
Sun Apr 3 10:54:44 UTC 2005


Issue status update for http://drupal.org/node/19298

 Project:      Drupal
 Version:      cvs
 Component:    base system
 Category:     feature requests
 Priority:     normal
 Assigned to:  Jeremy at kerneltrap.org
 Reported by:  Jeremy at kerneltrap.org
 Updated by:   Dries
 Status:       patch

Ber: personally, I would not involve user roles.  At least not until we
have a better understanding of how the loose chacing performs and
behaves.  Maybe in future, I'd add a _simple_ checkbox along these
lines: "Always use loose caching when the throttle mechanism is
active".
Daryl: could you try Jeremy's patch on SFX?  Ultimately a mechanism
like this ought to be tested on high-traffic websites.


Dries



Previous comments:
------------------------------------------------------------------------

March 23, 2005 - 06:32 : Jeremy at kerneltrap.org

Attachment: http://drupal.org/files/issues/database.mysql_1.patch (382 bytes)

Drupal's existing caching mechanism doesn't perform well on highly
dynamic websites in which the cache is flushed frequently.  One example
is a site that is under attack by a spambot that is posting spam
comments every few seconds, causing all cached pages to be flushed
every few seconds.
The attached patches provide the following cache methods:
 1) CACHE_DISABLED:  no caching.  This is unchanged from before.
 2) CACHE_ENABLED_STRICT: This is Drupal's existing caching method, in
which the cache is flushed immediately whenever it contains invalid
data.
 3) CACHE_ENABLED_LOOSE: This is the new caching method (feel free to
suggest a better name, this is the best I could come up with tonight). 
Essentially, it immediately flushes the cache only for specific users
who have modified cached data (whether or not they are logged in),
delaying the flushing of data for other users by several minutes.
This feature requires a new 'cache' field in the sessions table to
track the cache status for each user, whether or not they are logged
in.
A one line change to the system.module updates the cache configuration
settings to include an additional option for loose caching.
The rest of the changes are in bootstrap.inc, as follows:
 - cache_clear_all: if strict caching is enabled, nothing is changed. 
If loose caching is enabled, we delay the flushing of the entire cache
for 5 minutes.  Instead, we update this user's cache field in the
sessions table so we know any cache pages created before the current
time are no longer valid for this user.  We only set the global
"cache_flush" variable if it's not already set.  If the global
"cache_flush" variable was set more than 5 minutes ago, we flush all
cache pages that were expired when the "cache_flush" variable was set.
 - cache_get:  if strict caching is enabled, nothing is changed.  If
loose caching is enabled, we first check the global "cache_flush"
variable and if older than 5 minutes we do a garbage collection call to
cache_clear_all.  Next, we see if there is a valid cache entry for us. 
If the cache entry is older than our session table's "cache" field, we
don't use it.  This will cause the cache to be regenerated by the
current user, and ultimately updated for all users.
Attached is the patch to database.mysql.  To add manually, execute the
following:
ALTER TABLE sessions ADD cache int(11) NOT NULL default '0' AFTER
timestamp;


------------------------------------------------------------------------

March 23, 2005 - 06:33 : Jeremy at kerneltrap.org

Attachment: http://drupal.org/files/issues/system.module_5.patch (1.49 KB)

Attached is the patch for the system.module, adding the new cache
configuration option.


------------------------------------------------------------------------

March 23, 2005 - 06:36 : Jeremy at kerneltrap.org

Attachment: http://drupal.org/files/issues/bootstrap.inc.patch (3.62 KB)

And finally, the patch for bootstrap.inc.
BTW:  Do we want to make the 'cache_flush_delay' configurable?  Or
perhaps just stick with a sane default?  Perhaps 60 seconds is more
sane than 300 seconds?


------------------------------------------------------------------------

March 23, 2005 - 06:44 : Jeremy at kerneltrap.org

Attachment: http://drupal.org/files/issues/bootstrap.inc_0.patch (3.61 KB)

Updated version of the bootstrap.inc.patch, removing an erroneous "else"
that prevented the cache from being properly updated for a specific user
after a delayed flush.


------------------------------------------------------------------------

March 23, 2005 - 09:12 : Anonymous

Looks very good. Please provide patches in one file, though.
Gerhard


------------------------------------------------------------------------

March 23, 2005 - 10:35 : stefan nagtegaal

First of all, I do not have huge sites under my juristriction, so maybe
my points here aren't valid.. If so, tell me and I'll shut up..
I took a look at your patches, and I truly do not understand why you
set more Caching methods in 'admin/system'. I think that it would be
better if we stay with the cuurent options (Enabled, Disabled) and let
Drupal suggest the way of the caching methods (CACHE_ENABLED_LOOSE and
CACHE_ENABLED_SCTRICT). This would make your patch to system.module a
little smaller and easier to understand for newbies..
Also, as you said yourself, I don't see why we should make the
'cache_flush_delay' configurable. I mean, how would you know what a
good time would be? Couldn't drupal find out the right value for that,
so the administrator doesn't have to?
Again, I'm not speaking as someone who is in need of something like
this, but I can understand some people are.. What I'm trying to tell,
is that the idea is probably pretty good only with less options and
more automagical cache settings it would be more userfriendly ad
doesn't scare off newbies (or me)...


------------------------------------------------------------------------

March 23, 2005 - 12:21 : Dries

Looks like this patch can go in soon after the DRUPAL-4-6 has been
created.  You'll want to update the cache related documentation in
system_help() (then again it might not be necessary).  
I wouldn't bother making this more configurable until we have
evaluated/tested this some more.   I'm very interested in testing this
out or in getting some performance numbers.


------------------------------------------------------------------------

March 23, 2005 - 14:53 : Jeremy at kerneltrap.org

When time permits, I will combine the patches into one file, and update
the help text.  I included them as separate files for now to simplify
code review.  I also see a logical error I need to fix when performing
garbage collection that could cause the cache to flush uneccessarily.
I did not make the functionality automatic for two reasons:  it makes
the code slightly more complex, and it makes the functionality more
difficult to benchmark.  While I think it would be a good idea, I will
not modify the patch to implement automatic cache configuration unless
it is requested by Dries or other core developers.
To implement automatic cache configuration, it could simply be tied to
the throttle.  If the throttle is off (ie the throttle module is
disabled, or the site is not busy), use the strict cache.  If the
throttle is on (ie the throttle module is enabled and the site is
busy), use the loose cache.  
Alternatively, we could leave the configuration as it currently is in
this patch (with disabled, strict and loose), and the loose caching
functionality itself could be tied to the throttle.  If the throttle is
off, use a low cache_flush_delay (ie 15 seconds).  If the throttle is
on, use a high cache_flush_delay (ie 300 seconds).  The theory being
that the cache_flush_delay enforces a minimum cache life, and the
busier a site the more it will benefit from a longer minimum cache
life.  I believe that this latter implementation would be the best for
the spam floods I've experienced on my website.


------------------------------------------------------------------------

March 23, 2005 - 16:16 : moshe weitzman

changing the cache mode depending on throttle status is a really good
idea IMO. This patch is a great step forward for speeding up Drupal for
anon users. The biggest innovation in this area since bootstrap, I
think.


------------------------------------------------------------------------

March 25, 2005 - 14:30 : Jeremy at kerneltrap.org

Attachment: http://drupal.org/files/issues/cache.patch (8.82 KB)

Updated patch attached.  Fixed a couple small errors and put all patches
in one file.  I added a patch for database.pgsql and update.inc (though
the pgsql portion of update.inc may not be 100% right).
This patch does not tie the cache to the throttle, instead only
allowing manual configuration.


------------------------------------------------------------------------

April 1, 2005 - 15:19 : Jeremy at kerneltrap.org

Patch testimony:
I just got hit by a spam attack on KernelTrap this morning (more than
10,000 spam comment posted in less than 30 minutes) which choked my
server ("too many database connections").  So, I merged this patch into
my 4.5 installation and enabled "loose caching".  While I don't have any
actual performance numbers, I'm happy to report that a basically
non-functional website become comfortably responsive again.


------------------------------------------------------------------------

April 1, 2005 - 18:23 : Dries

The fact you use variable_get() in cache_get() presumably introduces an
extra SQL query for each page request that hits the cache (i.e. we have
to load and prepare all variables).  I haven't tested this though so
maybe we were already loading the variables.  Not too much of a
problem, I think, but I just wanted to point out this small
caveat/regression.
In the code comments, it might be a good idea to document the
interaction with the session-table.  Again, a minor nitpick.
This patch will go into Drupal 4.7.  Good work.


------------------------------------------------------------------------

April 1, 2005 - 20:06 : daryl at spreadfirefox.com

Chris Messina (factoryjoe) asked me to check this thread out and respond
based on my experience with www.spreadfirefox.com.
When we were really struggling with load issues and trying to make the
site scale, Dries suggested that we disable the automatic cache dump
per page load and handle cache dumping in a cron job; this removes all
per-request database overhead associated with cache dumping, and it
helped us out. Running such a cron job every five minutes or so beats
the heck out of running it for every request and would seem to achieve
roughly the same end that this patch achieves (at least as I read the
post -- I haven't looked at the actual cache). I suppose the downside
is that those without cron access are out of luck, and the patch I
gather sort of approximates a 5-minute cron routine dependent upon page
loads to run. Which seems fine to me.


------------------------------------------------------------------------

April 2, 2005 - 15:56 : Jeremy at kerneltrap.org

Dries wrote on April 1, 2005 - 11:23:
""The fact you use variable_get() in cache_get() presumably introduces
an extra SQL query for each page request that hits the cache"
"
I believe that this overhead is already assumed due to two uses of
variable_get() in drupal_page_header().  In building each page, we
first check 'dev_timer', and then 'cache'.  So my use of a variable
does not introduce an extra SQL query.
That said, it would be nice to get rid of this overhead.  Offhand, the
easiest way to do so would probably be to make both the 'dev_timer'
variable and the 'cache' variable configurable in settings.php instead
of from the administrative interface.  Why not?  It would be nice to
get rid of loading the variable table for cached pages...
Dries wrote on April 1, 2005 - 11:23:
""In the code comments, it might be a good idea to document the
interaction with the session-table."
"
Okay, I will update the patch.


------------------------------------------------------------------------

April 2, 2005 - 16:10 : Jeremy at kerneltrap.org

daryl at spreadfirefox.com wrote on April 1, 2005 - 13:06:
""Running such a cron job every five minutes or so beats the heck out
of running it for every request and would seem to achieve roughly the
same end that this patch achieves (at least as I read the post -- I
haven't looked at the actual cache). I suppose the downside is that
those without cron access are out of luck, and the patch I gather sort
of approximates a 5-minute cron routine dependent upon page loads to
run. Which seems fine to me."
"
Just to clarify, the patch does not approximate.  There is a similarity
to what you have described, however in the cron case the cache is dumped
every 5 minutes whether it needs it or not, and anyone who posts in that
5 minutes doesn't see the effect of their post until the cron run
completes.  
With this patch, the timer to dump the cache is only started once
something triggers a cache flush.  After the 5 minutes passes, all
cache entries older than 5 minutes will be dumped.  (This typically
happens the first time cache_get is called after 5 minutes has passed)
Additionally, the user who triggered the cache flush will actually stop
viewing cached pages so that any changes they made will show up for them
immediately in their session.  And as an added bonus, every page they
now visit will result in the cache for that page being rebuilt, helping
to further minimize the load by potentially spreading out over time the
cost of rebuilding cache entries.
Thus this has performance and usability advantages to simply using a
crontab entry.


------------------------------------------------------------------------

April 2, 2005 - 21:23 : moshe weitzman

i think you can simply get rid of the timer variable and assume it to be
true. it is harmless - certainly more harmless than its variable lookup!
as for the cache enable/disable, i agree that moving it to settings.php
is a good idea.


------------------------------------------------------------------------

April 3, 2005 - 06:20 : Jeremy at kerneltrap.org

Unfortunately my loose cache implementation currently relies on the
variable table, and I don't immediately see how to avoid this.  All the
configuration stuff can be either removed or moved into settings.php as
appropriate, but the 'cache_flush' variable is a dynamic timestamp that
needs to be shared with all sessions.  Currently this is done via the
variable table.
One idea is to move 'cache_flush' into the cache table...  It would
still cost a db_query, but certainly with less overhead than processing
the entire variable table.  Or is that too much of a kludge?  Thoughts?


------------------------------------------------------------------------

April 3, 2005 - 07:43 : Jeremy at kerneltrap.org

Attachment: http://drupal.org/files/issues/cache_0.patch (10.22 KB)

It's probably easier to see what I'm talking about with a patch.  So,
here's an updated patch that removes all variable_get's from
bootstrap.inc.  It works the same as my earlier loose caching patch,
but the cache configuration has been moved from the adminsitrative
interface to settings.php.  All the variable logic could probabily be
moved out of bootstrap.inc back to common.inc once this patch was
applied.
Specifically, this patch potentially improves upon my earlier patch by
removing the overhead of initializing the variable array.  I'd love to
see some benchmark results to determine if this is worth it.


------------------------------------------------------------------------

April 3, 2005 - 10:51 : Dries

Thanks for investigating this further, Jemery.  If variable_get() is
already called, I'd prefer to stick with it unless we pay a significant
performance penalty for it.


------------------------------------------------------------------------

April 3, 2005 - 12:34 : Bèr Kessels

Just a random thought I had lying around for a while:
Combine cache and throttle.
That would mean we use the cache we do now, but offer per-role cahce
refresh, based on the throttle level. Something like:
At Trottle limit [6     |v]
do not refresh for
[X] Anonymous
[X] Registered
[ ] Advertiser
[ ] Administrator
means that above level six, anonymous and registered cache will not be
refreshed.
This can be expanded by a cron wich will refres all cahces anyway.
meaning that in the above exmaple, witha cron every ten minutes,
anonymous and registered will see only refreshements every ten minutes.





More information about the drupal-devel mailing list