[Cache] Allow to configure serializator for cache instances. #27484

palex-fpt · 2018-06-04T06:05:34Z

Q	A
Branch?	master
Bug fix?	no
New feature?	yes
BC breaks?	no
Deprecations?	no
Tests pass?	-
Fixed tickets	-
License	MIT
Doc PR	-

Extract object serialization logic into SerializerInterface.
Allow opcache related caches to store __set_state enabled objects without serialization.

jvasseur · 2018-06-04T08:50:09Z

src/Symfony/Component/Cache/Serializer/NullSerializer.php

+
+use Symfony\Component\Cache\SerializerInterface;
+
+class NullSerializer implements SerializerInterface


Maybe this should be called IdentitySerializer, I expected a NullSerializer to always return null.

nicolas-grekas · 2018-06-04T09:42:24Z

src/Symfony/Component/Cache/Traits/PhpArrayTrait.php

-                } catch (\Exception $e) {
-                    throw new InvalidArgumentException(sprintf('Cache key "%s" has non-serializable array value.', $key), 0, $e);
+                    $exportedValue = $exportSerializer->serialize($value);
+                    $valuePart = 'array("isSerialized" => false, "value" => '.$exportedValue.')';


For me, an important criteria of this PR should be to preserve the original data format, so that we keep compatibility with existing dumped values. Here, this change it, and adds overhead to the storage, filling it with medadata that wasn't needed before.

I think the opcache-base storages should just not allow configuration of the serializer. WDYT?

Previous realization favors scalar and arrays of scalar data. Nulls and objects was stored in serialized form. It adds performance hit on every cache read or forces to use opcache for primitive types only.
It is possible to store values directly, but all checks in form: isset($this->values, $key) should be rewritten.

Previous realization favors scalar and arrays of scalar data. Nulls and objects was stored in serialized form

that's intended: only these benefit from opcache's shared memory

It adds performance hit on every cache read or forces to use opcache for primitive types only

do you have numbers about that to illustrate this statement?

Test: https://gist.github.com/palex-fpt/82fc3deed09c2de2a9023080a9613ce3

master results:
Duration (first/next) (largeArrayOfNulls): 14.200/2.300 s
Duration (first/next) (largeArrayOfStrings): 12.200/2.500 s
Duration (first/next) (largeArrayOfSmallObjects): 18.600/5.000 s
Duration (first/next) (largeObject): 130.400/112.300 s

branch results:
Duration (first/next) (largeArrayOfNulls): 16.500/3.500 s
Duration (first/next) (largeArrayOfStrings): 11.300/2.800 s
Duration (first/next) (largeArrayOfSmallObjects): 30.200/5.200 s
Duration (first/next) (largeObject): 145.700/3.400 s
Duration (first/next) (largeExportableObject): 7.500/3.300 s

nicolas-grekas · 2018-06-04T09:48:42Z

src/Symfony/Component/Cache/Traits/PhpArrayTrait.php

@@ -65,36 +68,25 @@ public function warmUp(array $values)

 EOF;

+        $exportSerializer = new PhpExportSerializer();


PhpExportSerializer is less capable than the current code, which accepts all serializable objects.
We should preserve this property.

PhpExportSerializer is backed by instance serializer, when data cannot be var_export-ed - it is serialized.

I can merge PhpExportSerializer into PhpArrayTrait. It is not used in other Caches and it is not part of serialization contract of PhpArrayTrait.

nicolas-grekas · 2018-06-04T09:49:43Z

src/Symfony/Component/Cache/Serializer/PhpSerializer.php

+        $unserializeCallbackHandler = ini_set('unserialize_callback_func', __CLASS__.'::handleUnserializeCallback');
+        try {
+            $value = unserialize($serialized);
+            if (false === $value && serialize(false) !== $serialized) {


this is a performance critical code, the previous logic should be preserved (checking the serialized "false" before unserializing)

It does unnecessary unserialize only on false value. Does storing false value as opcache entry is so frequent? But, ok. I would change it back.

nicolas-grekas · 2018-06-04T09:53:08Z

src/Symfony/Component/Cache/Traits/MemcachedTrait.php

@@ -206,14 +206,7 @@ protected function doSave(array $values, $lifetime)
     */
    protected function doFetch(array $ids)
    {
-        $unserializeCallbackHandler = ini_set('unserialize_callback_func', __CLASS__.'::handleUnserializeCallback');
-        try {
-            return $this->checkResultCode($this->getClient()->getMulti($ids));


serialization is also handled by the extension itself, so that the Memcached should be provided an IdentitySerializer and this could should be kept as is, isn't it?

I did not add SerilaizerTrait to Caches that handle serialization by itselves: Memcached, Apcu, Doctrine. It looks like adding IdentitySerilizer to it is good idea.

it would allow using a compressing serializer if needed, so would be great yes

palex-fpt · 2018-06-07T01:32:27Z

Add test result at https://gist.github.com/palex-fpt/82fc3deed09c2de2a9023080a9613ce3

nicolas-grekas · 2018-06-07T06:19:13Z

I just realized we cannot use the var_export() strategy with PhpArrayAdapter, because this strategy requires instantiating all objects in the pool, when typically only a few are needed per run.

nicolas-grekas · 2018-06-07T06:20:33Z

For PhpFileAdapter, var_export() poses another issue, which is deep cloning, which is broken. :(

palex-fpt · 2018-06-07T06:28:50Z

Was PhpArrayAdapter designed with that access strategy in mind?
PhpArrayAdapter - fast access to many cache entries. But heavy on memory usage.
PhpFilesAdapter - fast access to single cache entry. But looses to PhpArrayAdapter on multiple entries.

palex-fpt · 2018-06-07T06:32:33Z

Object should be tested against serialization method. There is no universal serializator.

nicolas-grekas · 2018-06-07T06:32:34Z

Oh yes it was. PhpArray is basically free on memory usage, because opcache puts the structure in shared memory, which is out if quota.

nicolas-grekas · 2018-06-07T06:34:00Z

See eg. https://blog.blackfire.io/speeding-up-autoloading-on-php-5-6-7-0-for-everyone.html

nicolas-grekas · 2018-06-07T06:37:42Z

My recommendation for this PR would be to remove the export-based serializers and leave the PHP ones as is. That would make a first important step mergeable. Then, if we can find a better serialization strategy, let's do it in another PR.

palex-fpt · 2018-06-07T06:45:13Z

So, for PhpArrayCache we want to store scalar/serialized data and unserialize it on demand.
Does same strategy required for PhpFilesAdapter?

nicolas-grekas · 2018-06-07T06:53:14Z

For PhpArray, we want to create the objects on-demand yes.
For PhpFile also, but this is already achieved by the file container. Still there is another thing we need: deep copying of potentially nested objects.

palex-fpt · 2018-06-07T06:57:26Z

PhpFile uses same strategy as PhpArray. It stores non-scalar values in serialized form. That add un-serialization performance hit on access.

palex-fpt · 2018-06-07T10:29:56Z

Main point of this PR was to speed up access to opcache entries. This can be achieved by removing serialization part (allow to var_export objects with __set_state) or by changing serializator (igbinary ex.). It is possible to add extension points to select types that would be var_exported or to select serializator. But it is not possible to do that in PhpFilesCache preserving current data format. PhpFilesCache detects types by inspecting serialized content. Should it be done in fresh new implementation of CacheInterface? Can we change current data format?

nicolas-grekas · 2018-06-07T15:07:52Z

Main point of this PR was to speed up access to opcache entries

I had a different goal, which you already achieved here (having a swappable serialization format, of special interest when using remote backends IMHO, to allow using e.g. igbinary.)

For OPcache entries, I propose #27543 instead.

palex-fpt · 2018-06-08T00:11:06Z

I had a different goal, which you already achieved here (having a swappable serialization format, of special interest when using remote backends IMHO, to allow using e.g. igbinary.)

Ok. I would cleanup all changes from PhpFilesCache and PhpArrayCache.

IMHO we don't need IdentitySerializers in Apcu, Doctrine, Memcached, Redis.
There are caches that use binary data backend (Filesystem, Pdo). That caches should be supplied with Serializers. And there are caches that has native (or through extension) php support (Acpu, Memcached, Redis, Doctrine). There is no need to add IdentitySerializers to its.
In case someone would like to transform data before sending it to backend cache he can use ProxyAdapter to change data type.

nicolas-grekas · 2018-06-18T07:48:19Z

How should we move this forward? Taking #27543 into account, I would suggest reverting all changes to Php* and Array* from this PR (moving the patch to a later PR maybe), and focus on providing alternative serializers to adapter/caches that extend the AbstractCache / AbstractAdapter. This would allow using igbinary/whatever where it's most useful.
About naming, it's suggest using marshall again, so that there is no possible confusion with the Serializer component. I'm sorry about the flip flap here, I know I asked for the opposite before.
I moved #27543 in the Marshaller namespace, so that this PR could use the same (the PhpSerializer class should keep its name, to not collide with PhpMarshaller from #27543.)
WDYT?

@palex-fpt

…sible (nicolas-grekas) This PR was merged into the 4.2-dev branch. Discussion ---------- [Cache] serialize objects using native arrays when possible | Q | A | ------------- | --- | Branch? | master | Bug fix? | no | New feature? | no | BC breaks? | no | Deprecations? | no | Tests pass? | yes | Fixed tickets | - | License | MIT | Doc PR | - This PR allows leveraging OPCache shared memory when storing objects in `Php*` pool storages (as done by default for all system caches). This improves performance a bit further when loading e.g. annotations, etc. (bench coming); Instead of using native php serialization, this uses a marshaller that represents objects in plain static arrays. Unmarshalling these arrays is faster than unserializing the corresponding PHP strings (because it works with copy-on-write, while unserialize cannot.) php-serialization is still a possible format because we have to use it when serializing structures with internal references or with objects implementing `Serializable`. The best serialization format is selected automatically so this is completely seamless. ping @palex-fpt since you gave me the push to work on this, and are pursuing a similar goal in #27484. I'd be thrilled to get some benchmarks on your scenarios. Commits ------- 866420e [Cache] serialize objects using native arrays when possible

nicolas-grekas · 2018-06-18T16:24:03Z

#27543 is now merged, this can be rebased :)

palex-fpt · 2018-06-19T05:31:51Z

Php* changes was already reverted. After reverting Array* it left only two caches: FilesystemCache and PdoCache. My end goal is to use igbinary for filesystem cache. Can we just add some boolean flag to Filesystem* constructor to use igbinary?

Tobion · 2018-06-19T08:04:36Z

src/Symfony/Component/Cache/Traits/SerializerTrait.php

+    /** @var SerializerInterface */
+    private $serializer;
+
+    public function setSerializer(SerializerInterface $serializer)


does this need to be public? it's currently not used in a public way.

nicolas-grekas · 2018-06-20T06:45:35Z

I closing in favor of #27645, which provides auto-adaptative igbinary support. It does not provide serializer injection, but I'm not sure we need any actually. I would like to thank you for providing this PR anyway, and for your comments on the other cache-related PRs, it's been helping A LOT to improve the component. Let's continue on #27645.

carsonbot added Status: Needs Review Cache Feature labels Jun 4, 2018

palex-fpt force-pushed the cache-marshaller branch 2 times, most recently from 9a2b820 to 2e76ae4 Compare June 4, 2018 06:29

palex-fpt mentioned this pull request Jun 4, 2018

[Cache] Use igbinary when available #19895

Closed

jvasseur reviewed Jun 4, 2018

View reviewed changes

nicolas-grekas reviewed Jun 4, 2018

View reviewed changes

nicolas-grekas added this to the next milestone Jun 4, 2018

palex-fpt force-pushed the cache-marshaller branch from 26ba1cf to a4b045f Compare June 7, 2018 06:21

palex-fpt force-pushed the cache-marshaller branch from a4b045f to 6401ce7 Compare June 7, 2018 06:37

palex-fpt force-pushed the cache-marshaller branch from 6401ce7 to defdc50 Compare June 7, 2018 06:39

nicolas-grekas mentioned this pull request Jun 7, 2018

[Cache] serialize objects using native arrays when possible #27543

Merged

palex-fpt force-pushed the cache-marshaller branch 3 times, most recently from f1b3fe3 to b279460 Compare June 9, 2018 04:45

cache-serializer

9a94e6c

palex-fpt force-pushed the cache-marshaller branch from b279460 to 9a94e6c Compare June 14, 2018 04:37

Tobion reviewed Jun 19, 2018

View reviewed changes

nicolas-grekas closed this Jun 20, 2018

nicolas-grekas modified the milestones: next, 4.2 Nov 1, 2018

yellow1912 mentioned this pull request Dec 7, 2020

[Cache] Allow skipping serializer for a specific cache pool #39355

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Cache] Allow to configure serializator for cache instances. #27484

[Cache] Allow to configure serializator for cache instances. #27484

palex-fpt commented Jun 4, 2018

jvasseur Jun 4, 2018

nicolas-grekas Jun 4, 2018

palex-fpt Jun 4, 2018

nicolas-grekas Jun 4, 2018

palex-fpt Jun 5, 2018

nicolas-grekas Jun 4, 2018

palex-fpt Jun 4, 2018

palex-fpt Jun 4, 2018

nicolas-grekas Jun 4, 2018

palex-fpt Jun 4, 2018

nicolas-grekas Jun 4, 2018

palex-fpt Jun 4, 2018

nicolas-grekas Jun 4, 2018

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

palex-fpt commented Jun 7, 2018

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018 •

edited

Loading

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018 •

edited

Loading

palex-fpt commented Jun 7, 2018

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018 •

edited

Loading

palex-fpt commented Jun 8, 2018

nicolas-grekas commented Jun 18, 2018

nicolas-grekas commented Jun 18, 2018

palex-fpt commented Jun 19, 2018

Tobion Jun 19, 2018

nicolas-grekas commented Jun 20, 2018 •

edited

Loading


		use Symfony\Component\Cache\SerializerInterface;

		class NullSerializer implements SerializerInterface

		@@ -65,36 +68,25 @@ public function warmUp(array $values)

		EOF;

		$exportSerializer = new PhpExportSerializer();

[Cache] Allow to configure serializator for cache instances. #27484

[Cache] Allow to configure serializator for cache instances. #27484

Conversation

palex-fpt commented Jun 4, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

palex-fpt commented Jun 7, 2018

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018 • edited Loading

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018 • edited Loading

palex-fpt commented Jun 7, 2018

palex-fpt commented Jun 7, 2018

nicolas-grekas commented Jun 7, 2018 • edited Loading

palex-fpt commented Jun 8, 2018

nicolas-grekas commented Jun 18, 2018

nicolas-grekas commented Jun 18, 2018

palex-fpt commented Jun 19, 2018

Choose a reason for hiding this comment

nicolas-grekas commented Jun 20, 2018 • edited Loading

nicolas-grekas commented Jun 7, 2018 •

edited

Loading

nicolas-grekas commented Jun 7, 2018 •

edited

Loading

nicolas-grekas commented Jun 7, 2018 •

edited

Loading

nicolas-grekas commented Jun 20, 2018 •

edited

Loading