[HttpFoundation] Add StreamedJsonResponse for efficient JSON streaming

[alexander-schranz]

Q	A
Branch?	6.2
Bug fix?	no
New feature?	yes
Deprecations?	no
Tickets	Fix #...
License	MIT
Doc PR	symfony/symfony-docs#17301

When big data are streamed via JSON API it can sometimes be difficult to keep the resources usages low. For this I experimented with a different way of streaming data for JSON responses. It uses combination of structured array and generics which did result in a lot better result.

More can be read about here: https://github.com/alexander-schranz/efficient-json-streaming-with-symfony-doctrine.

I thought it maybe can be a great addition to Symfony itself to make this kind of responses easier and that APIs can be made more performant.

Usage

First Version (replaced)

class ArticleListAction {
    public function __invoke(EntityManagerInterface  $entityManager): Response
    {
        $articles = $this->findArticles($entityManager);

        return new StreamedJsonResponse(
            // json structure with replacers identifiers
            [
                '_embedded' => [
                    'articles' => '__articles__',
                ],
            ],
            // array of generator replacer identifier used as key
            [
                '__articles__' => $this->findArticles('Article'),
            ]
        );
    }

    private function findArticles(EntityManagerInterface  $entityManager): \Generator
    {
        $queryBuilder = $entityManager->createQueryBuilder();
        $queryBuilder->from(Article::class, 'article');
        $queryBuilder->select('article.id')
            ->addSelect('article.title')
            ->addSelect('article.description');

        return $queryBuilder->getQuery()->toIterable();
    }
}

Update Version (thx to @ro0NL for the idea):

class ArticleListAction {
    public function __invoke(EntityManagerInterface  $entityManager): Response
    {
        $articles = $this->findArticles($entityManager);

        return new StreamedJsonResponse(
            // json structure with generators in it which are streamed
            [
                '_embedded' => [
                    'articles' => $this->findArticles('Article'), // returns a generator which is streamed
                ],
            ],
        );
    }

    private function findArticles(EntityManagerInterface  $entityManager): \Generator
    {
        $queryBuilder = $entityManager->createQueryBuilder();
        $queryBuilder->from(Article::class, 'article');
        $queryBuilder->select('article.id')
            ->addSelect('article.title')
            ->addSelect('article.description');

        return $queryBuilder->getQuery()->toIterable();
    }
}

---

As proposed by @OskarStark the Full Content of Blog about "Efficient JSON Streaming with Symfony and Doctrine":

Efficient JSON Streaming with Symfony and Doctrine

After reading a tweet about we provide only a few items (max. 100) over our
JSON APIs but providing 4k images for our websites. I did think about why is
this the case.

The main difference first we need to know about how images are streamed.
On webservers today is mostly the sendfile feature used. Which is very
efficient as it can stream a file chunk by chunk and don't need to load
the whole data.

So I'm asking myself how we can achieve the same mechanisms for our
JSON APIs, with a little experiment.

As an example we will have a look at a basic entity which has the
following fields defined:

• id: int
• title: string
• description: text

The response of our API should look like the following:

{
  "_embedded": {
    "articles": [
      {
        "id": 1,
        "title": "Article 1",
        "description": "Description 1\nMore description text ...",
      },
      ...
    ]
  } 
}

Normally to provide this API we would do something like this:

<?php

namespace App\Controller;

use App\Entity\Article;
use Doctrine\ORM\EntityManagerInterface;
use Symfony\Component\HttpFoundation\JsonResponse;
use Symfony\Component\HttpFoundation\Response;

class ArticleListAction
{
    public function __invoke(EntityManagerInterface $entityManager): Response
    {
        $articles = $this->findArticles($entityManager);

        return JsonResponse::fromJsonString(json_encode([
            'embedded' => [
                'articles' => $articles,
            ],
            'total' => 100_000,
        ], JSON_THROW_ON_ERROR | JSON_UNESCAPED_SLASHES | JSON_UNESCAPED_UNICODE));
    }

    // normally this method would live in a repository
    private function findArticles(EntityManagerInterface  $entityManager): iterable
    {
        $queryBuilder = $entityManager->createQueryBuilder();
        $queryBuilder->from(Article::class, 'article');
        $queryBuilder->select('article.id')
            ->addSelect('article.title')
            ->addSelect('article.description');

        return $queryBuilder->getQuery()->getResult();
    }
}

In most cases we will add some pagination to the endpoint so our response are not too big.

Making the api more efficient

But there is also a way how we can stream this response in an efficient way.

First of all we need to adjust how we load the articles. This can be done by replace
the getResult with the more efficient toIterable:

-        return $queryBuilder->getQuery()->getResult();
+        return $queryBuilder->getQuery()->toIterable();

Still the whole JSON need to be in the memory to send it. So we need also refactoring
how we are creating our response. We will replace our JsonResponse with the
StreamedResponse object.

return new StreamedResponse(function() use ($articles) {
    // stream json
}, 200, ['Content-Type' => 'application/json']);

But the json format is not the best format for streaming, so we need to add some hacks
so we can make it streamable.

First we will create will define the basic structure of our JSON this way:

$jsonStructure = json_encode([
    'embedded' => [
        'articles' => ['__REPLACES_ARTICLES__'],
    ],
    'total' => 100_000,
], JSON_THROW_ON_ERROR | JSON_UNESCAPED_SLASHES | JSON_UNESCAPED_UNICODE);

Instead of the $articles we are using a placeholder which we use to split the string into
a $before and $after variable:

[$before, $after] = explode('"__REPLACES_ARTICLES__"', $jsonStructure, 2);

Now we are first sending the $before:

echo $before . PHP_EOL;

Then we stream the articles one by one to it here we need to keep the comma in mind which
we need to add after every article but not the last one:

foreach ($articles as $count => $article) {
    if ($count !== 0) {
        echo ',' . PHP_EOL; // if not first element we need a separator
    }

    echo json_encode($article, JSON_THROW_ON_ERROR | JSON_UNESCAPED_SLASHES | JSON_UNESCAPED_UNICODE);
}

Also we will add an additional flush after every 500 elements:

if ($count % 500 === 0 && $count !== 100_000) { // flush response after every 500
    flush();
}

After that we will also send the $after part:

echo PHP_EOL;
echo $after;

The result

So at the end the whole action looks like the following:

<?php

namespace App\Controller;

use App\Entity\Article;
use Doctrine\ORM\EntityManagerInterface;
use Symfony\Component\HttpFoundation\Response;
use Symfony\Component\HttpFoundation\StreamedResponse;

class ArticleListAction
{
    public function __invoke(EntityManagerInterface  $entityManager): Response
    {
        $articles = $this->findArticles($entityManager);

        return new StreamedResponse(function() use ($articles) {
            // defining our json structure but replaces the articles with a placeholder
            $jsonStructure = json_encode([
                'embedded' => [
                    'articles' => ['__REPLACES_ARTICLES__'],
                ],
                'total' => 100_000,
            ], JSON_THROW_ON_ERROR | JSON_UNESCAPED_SLASHES | JSON_UNESCAPED_UNICODE);

            // split by placeholder
            [$before, $after] = explode('"__REPLACES_ARTICLES__"', $jsonStructure, 2);

            // send first before part of the json
            echo $before . PHP_EOL;

            // stream article one by one as own json
            foreach ($articles as $count => $article) {
                if ($count !== 0) {
                    echo ',' . PHP_EOL; // if not first element we need a separator
                }

                if ($count % 500 === 0 && $count !== 100_000) { // flush response after every 500
                    flush();
                }

                echo json_encode($article, JSON_THROW_ON_ERROR | JSON_UNESCAPED_SLASHES | JSON_UNESCAPED_UNICODE);
            }

            // send the after part of the json as last
            echo PHP_EOL;
            echo $after;
        }, 200, ['Content-Type' => 'application/json']);
    }

    private function findArticles(EntityManagerInterface  $entityManager): iterable
    {
        $queryBuilder = $entityManager->createQueryBuilder();
        $queryBuilder->from(Article::class, 'article');
        $queryBuilder->select('article.id')
            ->addSelect('article.title')
            ->addSelect('article.description');

        return $queryBuilder->getQuery()->toIterable();
    }
}

The metrics for 100000 Articles (nginx + php-fpm 7.4 - Macbook Pro 2013):

	Old Implementation	New Implementation
Memory Usage	49.53 MB	2.10 MB
Memory Usage Peak	59.21 MB	2.10 MB
Time to first Byte	478ms	28ms
Time	2.335 s	0.584 s

This way we did not only reduce the memory usage on our server
also we did make the response faster. The memory usage was
measured here with memory_get_usage and memory_get_peak_usage.
The "Time to first Byte" by the browser value and response times
over curl.

Updated 2022-10-02 - (symfony serve + php-fpm 8.1 - Macbook Pro 2021)

	Old Implementation	New Implementation
Memory Usage	64.21 MB	2.10 MB
Memory Usage Peak	73.89 MB	2.10 MB
Time to first Byte	0.203 s	0.049 s
Updated Time (2022-10-02)	0.233 s	0.232 s

While there is not much different for a single response in the time,
the real performance is the lower memory usage. Which will kick in when
you have a lot of simultaneously requests. On my machine >150 simultaneously
requests - which is a high value but will on a normal server be a lot lower.

While 150 simultaneously requests crashes in the old implementation
the new implementation still works with 220 simultaneously requests. Which
means we got about ~46% more requests possible.

Reading Data in javascript

As we stream the data we should also make our JavaScript on the other
end the same way - so data need to read in streamed way.

Here I'm just following the example from the Fetch API Processing a text file line by line

So if we look at our script.js we split the object
line by line and append it to our table. This method is definitely not the
way how JSON should be read and parsed. It should only be shown as example
how the response could be read from a stream.

Conclusion

The implementation looks a little hacky for maintainability it could
be moved into its own Factory which creates this kind of response.

Example:

return StreamedResponseFactory::create(
    [
        'embedded' => [
            'articles' => ['__REPLACES_ARTICLES__'],
        ],
        'total' => 100_000,
    ],
    ['____REPLACES_ARTICLES__' => $articles]
);

The JavaScript part something is definitely not ready for production
and if used you should probably creating your own content-type e.g.:
application/json+stream. So you are parsing the json this way
only when you know it is really in this line by line format.
There maybe better libraries like JSONStream
to read data but at current state did test them out. Let me know
if somebody has experience with that and has solutions for it.

Atleast what I think everybody should use for providing lists
is to use toIterable when possible for your lists when loading
your data via Doctrine and and select specific fields instead
of using the ORM to avoid hydration process to object.

Let me know what you think about this experiment and how you currently are
providing your JSON data.

The whole experiment here can be checked out and test yourself via this repository.

Attend the discussion about this on Twitter.

Update 2022-09-27

Added a StreamedJsonRepsonse class and
try to contribute this implementation to the Symfony core.

#47709

Update 2022-10-02

Updated some statistics with new machine and apache benchmark tests for concurrency requests.

[alexander-schranz]

The error in the tests of Stopwatch is unrelated to the pull request.

[ro0NL]

would it be reasonable to consider a "compute json inline" approach, rather than end-users taking care of unique identifiers

$lazyJson = ['key' => fn() => yield from $heavy];

[stof]

@ro0NL this would force to re-implement the whole json encoding in userland

[ro0NL]

we could array walk the structure first, thus keeping the unique placeholders an implementation detail.

[stof]

@ro0NL if you do that, you are not streaming json anymore, defeating the whole purpose of this PR.

[ro0NL]

the idea is to split the generators from the structure, preserving remaining logic. But this is an extra step yes, thus less ideal perhaps.

[alexander-schranz]

@ro0NL interesting input. As I think the structure array is mostly small it could be possible. But we would need to have a look at what difference this would be in the performance.

I hacked something together using array_walk_recursive: https://3v4l.org/tndhO. Will have a deeper look at it at the evening or next days.

[stof]

@alexander-schranz be careful when implementing this. is_callable would turn some strings into placeholders instead of outputting them.

[alexander-schranz]

@stof great hint think $item instanceof Closure should then do the job?

[stof]

now that we have first class callables, I would say yes. You can convert any callable to a closure using this feature.

[alexander-schranz]

Okay I don't need to check for closures or callables. I just need to check on \Generators because the Closures are already called. Which is very important, as example when Database connection is not available the exception need to be thrown in the Controller and should not be thrown when Status Code 200 is already returned:

return new StreamedJsonResponse(
    [
        '_embedded' => [
            'articles' => $this->findArticles('Article'), // returns a \Generator which will generate a list of data
        ],
    ],
);

The diff between old and new implementation is not big it just takes about 0.0000128s todo the array_walk_recursive and replace it. It also did not have any visible changes on the memory usage. The tested arrays are really small but that will mostly be the case I think in this kind of responses.

I also update the example repository using the new class under /symfony-articles.json: https://github.com/alexander-schranz/efficient-json-streaming-with-symfony-doctrine if somebody want to experiment with it.

[OskarStark]

I propose to add the content from the README of your prototype application to the PR header 👍🏻

[alexander-schranz]

@OskarStark added.

Think PR is blocked until 6.3 branch is created?

[OskarStark]

@OskarStark added.

thanks

Think PR is blocked until 6.3 branch is created?

Yes

[dunglas]

For the record, @mtarld @soyuka and I are working on a new component that will be an alternative to json_encode/json_decode and to the Symfony Serializer that will natively support JSON streaming (for encoding and decoding). Maybe will it be possible to use this component in this PR.

[alexander-schranz]

@dunglas that sounds very interesting. I think currently I would stay with the implementation how it is for now, this gives a very low resource solution without the need that the http foundation package has additional requirements to any kind of serializer and so on. Still a serializer/normalizer is possible be used inside the Generator already, which will be they current implementation of this class also be very low on resources usage as it don't try to serialize all objects at once just one after the other and so don't need to keep more then one object in the memory aslong as the ORM loading allows that.

[chalasr]

Shall we move forward on this one?

[alexander-schranz]

@chalasr rebased. Not sure what is open or required to get this merged :)

[chalasr]

Let's iterate, thanks @alexander-schranz!

[alexander-schranz]

🎉 Thx you all for the great feedback and ideas. Think we got a great solution out of it with a better DX as I could think of when created the Pull request.

@chalasr that sounds great :)