Peterbe.com

This web app I'm working on gets a blob of bytes from a HTTP POST. The nature of the blob is a 100MB to 1,100MB blob of a zip file. What my app currently does is that it takes this byte buffer, uses Python's built in zipfile to extract all its content to a temporary directory. A second function then loops over the files within this extracted tree and processes each file in multiple threads with concurrent.futures.ThreadPoolExecutor. Here's the core function itself:

@metrics.timer_decorator('upload_dump_and_extract')
def dump_and_extract(root_dir, file_buffer):
    zf = zipfile.ZipFile(file_buffer)
    zf.extractall(root_dir)

So far so good.

Speed Speed Speed

I quickly noticed that this is amounting to quite a lot of time spent doing the unzip and the writing to disk. What to do????

At first I thought I'd shell out to good old unzip. E.g. unzip -d /tmp/tempdirextract /tmp/input.zip but that has two flaws:

1) I'd first have to dump the blob of bytes to disk and do the overhead of shelling out (i.e. Python subprocess)
2) It's actually not faster. Did some experimenting and got the same results at Alex Martelli in this Stackoverflow post

What about disk speed? Yeah, this is likely to be a part of the total time. The servers that run the symbols.mozilla.org service runs on AWS EC2 c4.large. This only has EBS (Elastic Block Storage). However, AWS EC2 c3.large looks interesting since it's using SSD disks. That's probably a lot faster. Right?

Note! For context, the kind of .zip files I'm dealing with contain many small files and often 1-2 really large ones.

EC2s Benchmarking

I create two EC2 nodes to experiment on. One c3.large and one c4.large. Both running Ubuntu 16.04.

Next, I have this little benchmarking script which loops over a directory full of .zip files between 200MB-600MB large. Roughly 10 of them. It then loads each one, one at a time, into memory and calls the dump_and_extract. Let's run it on each EC2 instance:

On c4.large

c4.large$ python3 fastest-dumper.py /tmp/massive-symbol-zips
138.2MB/s            291.1MB              2.107s
146.8MB/s            314.5MB              2.142s
144.8MB/s            288.2MB              1.990s
84.5MB/s             532.4MB              6.302s
146.6MB/s            314.2MB              2.144s
136.5MB/s            270.7MB              1.984s
85.9MB/s             518.9MB              6.041s
145.2MB/s            306.8MB              2.113s
127.8MB/s            138.7MB              1.085s
107.3MB/s            454.8MB              4.239s
141.6MB/s            251.2MB              1.774s


Average speed: 127.7MB/s
Median speed:  138.2MB/s

Average files created:       165
Average directories created: 129

On c3.large

c3.large$ python3 fastest-dumper.py -t /mnt/extracthere /tmp/massive-symbol-zips
105.4MB/s            290.9MB              2.761s
98.1MB/s             518.5MB              5.287s
108.1MB/s            251.2MB              2.324s
112.5MB/s            294.3MB              2.615s
113.7MB/s            314.5MB              2.767s
106.3MB/s            291.5MB              2.742s
104.8MB/s            291.1MB              2.778s
114.6MB/s            248.3MB              2.166s
114.2MB/s            248.2MB              2.173s
105.6MB/s            298.1MB              2.823s
106.2MB/s            297.6MB              2.801s
98.6MB/s             521.4MB              5.289s


Average speed: 107.3MB/s
Median speed:  106.3MB/s

Average files created:       165
Average directories created: 127

What the heck!? The SSD based instance is 23% slower!

I ran it a bunch of times and the average and median numbers are steady. c4.large is faster than c3.large at unzipping large blobs to disk. So much for that SSD!

Something Weird Is Going On

It's highly likely that the unzipping work is CPU bound and that most of those, for example, 5 seconds is spent unzipping and only a small margin is the time it takes to write to disk.

If the unzipping CPU work is the dominant "time consumer" why is there a difference at all?!

Or, is the "compute power" the difference between c3 and c4 and disk writes immaterial?

For the record, this test clearly demonstrates that the locally mounted SSD drive is 600% faster than ESB.

c3.large$ dd if=/dev/zero of=/tmp/1gbtest bs=16k count=65536
65536+0 records in
65536+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 16.093 s, 66.7 MB/s
c3.large$ sudo dd if=/dev/zero of=/mnt/1gbtest bs=16k count=65536
65536+0 records in
65536+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 2.62728 s, 409 MB/s

Let's try again. But instead of using c4.large and c3.large, let's use the beefier c4.4xlarge and c3.4xlarge. Both have 16 vCPUs.

c4.4xlarge

c4.4xlarge$ python3 fastest-dumper.py /tmp/massive-symbol-zips
130.6MB/s            553.6MB              4.238s
149.2MB/s            297.0MB              1.991s
129.1MB/s            529.8MB              4.103s
116.8MB/s            407.1MB              3.486s
147.3MB/s            306.1MB              2.077s
151.9MB/s            248.2MB              1.634s
140.8MB/s            292.3MB              2.076s
146.8MB/s            288.0MB              1.961s
142.2MB/s            321.0MB              2.257s


Average speed: 139.4MB/s
Median speed:  142.2MB/s

Average files created:       148
Average directories created: 117

c3.4xlarge

c3.4xlarge$ python3 fastest-dumper.py -t /mnt/extracthere /tmp/massive-symbol-zips
95.1MB/s             502.4MB              5.285s
104.1MB/s            303.5MB              2.916s
115.5MB/s            313.9MB              2.718s
105.5MB/s            517.4MB              4.904s
114.1MB/s            288.1MB              2.526s
103.3MB/s            555.9MB              5.383s
114.0MB/s            288.0MB              2.526s
109.2MB/s            251.2MB              2.300s
108.0MB/s            291.0MB              2.693s


Average speed: 107.6MB/s
Median speed:  108.0MB/s

Average files created:       150
Average directories created: 119

What's going on!? The time it takes to unzip and write to disk is, on average, the same for c3.large as c3.4xlarge!

Is Go Any Faster?

I need a break. As mentioned above, the unzip command line program is not any better than doing it in Python. But Go is faster right? Right?

Please first accept that I'm not a Go programmer even though I can use it to build stuff but really my experience level is quite shallow.

Here's the Go version. Critical function that does the unzipping and extraction to disk here:

func DumpAndExtract(dest string, buffer []byte, name string) {
    size := int64(len(buffer))
    zipReader, err := zip.NewReader(bytes.NewReader(buffer), size)
    if err != nil {
        log.Fatal(err)
    }
    for _, f := range zipReader.File {
        rc, err := f.Open()
        if err != nil {
            log.Fatal(err)
        }
        defer rc.Close()
        fpath := filepath.Join(dest, f.Name)
        if f.FileInfo().IsDir() {
            os.MkdirAll(fpath, os.ModePerm)
        } else {
            // Make File
            var fdir string
            if lastIndex := strings.LastIndex(fpath, string(os.PathSeparator)); lastIndex > -1 {
                fdir = fpath[:lastIndex]
            }
            err = os.MkdirAll(fdir, os.ModePerm)
            if err != nil {
                log.Fatal(err)
            }
            f, err := os.OpenFile(
                fpath, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, f.Mode())
            if err != nil {
                log.Fatal(err)
            }
            defer f.Close()

            _, err = io.Copy(f, rc)
            if err != nil {
                log.Fatal(err)
            }
        }
    }
}

And the measurement is done like this:

size := int64(len(content))
t0 := time.Now()
DumpAndExtract(tmpdir, content, filename)
t1 := time.Now()
speed := float64(size) / t1.Sub(t0).Seconds()

It's not as sophisticated (since it's only able to use /tmp) but let's just run it see how it compares to Python:

c4.4xlarge$ mkdir ~/GO
c4.4xlarge$ export GOPATH=~/GO
c4.4xlarge$ go get github.com/pyk/byten
c4.4xlarge$ go build unzips.go
c4.4xlarge$ ./unzips /tmp/massive-symbol-zips
56MB/s         407MB          7.27804954
74MB/s         321MB          4.311504933
75MB/s         288MB          3.856798853
75MB/s         292MB          3.90972474
81MB/s         248MB          3.052652168
58MB/s         530MB          9.065985117
59MB/s         554MB          9.35237202
75MB/s         297MB          3.943132388
74MB/s         306MB          4.147176578

Average speed:    70MB/s
Median speed:     81MB/s

So... Go is, on average, 40% slower than Python in this scenario. Did not expect that.

In Conclusion

No conclusion. Only confusion.

I thought this would be a lot clearer and more obvious. Yeah, I know it's crazy to measure two things at the same time (unzip and disk write) but the whole thing started with a very realistic problem that I'm trying to solve. The ultimate question was; will the performance benefit from us moving the web servers from AWS EC2 c4.large to c3.large and I think the answer is no.

UPDATE (Nov 30, 2017)

Here's a horrible hack that causes the extraction to always go to /dev/null:

class DevNullZipFile(zipfile.ZipFile):
    def _extract_member(self, member, targetpath, pwd):
        # member.is_dir() only works in Python 3.6
        if member.filename[-1] == '/':
            return targetpath
        dest = '/dev/null'
        with self.open(member, pwd=pwd) as source, open(dest, "wb") as target:
            shutil.copyfileobj(source, target)
        return targetpath


def dump_and_extract(root_dir, file_buffer, klass):
    zf = klass(file_buffer)
    zf.extractall(root_dir)

And here's the outcome of running that:

c4.4xlarge$ python3 fastest-dumper.py --dev-null /tmp/massive-symbol-zips
170.1MB/s            297.0MB              1.746s
168.6MB/s            306.1MB              1.815s
147.1MB/s            553.6MB              3.765s
132.1MB/s            407.1MB              3.083s
145.6MB/s            529.8MB              3.639s
175.4MB/s            248.2MB              1.415s
163.3MB/s            321.0MB              1.965s
162.1MB/s            292.3MB              1.803s
168.5MB/s            288.0MB              1.709s


Average speed: 159.2MB/s
Median speed:  163.3MB/s

Average files created:       0
Average directories created: 0

I ran it a few times to make sure the numbers are stable. They are. This is on the c4.4xlarge.

So, the improvement of writing to /dev/null instead of the ESB /tmp is 15%. Kinda goes to show how much of the total time is spent reading the ZipInfo file object.

For the record, the same comparison on the c3.4xlarge was 30% improvement when using /dev/null.

Also for the record, if I replace that line shutil.copyfileobj(source, target) above with pass, the average speed goes from 159.2MB/s to 112.8GB/s but that's not a real value of any kind.

UPDATE (Nov 30, 2017)

Here's the same benchmark using c5.4xlarge instead. So, still EBS but...
"3.0 GHz Intel Xeon Platinum processors with new Intel Advanced Vector Extension 512 (AVX-512) instruction set"

Let's run it on this supposedly faster CPU:

c5.4xlarge$ python3 fastest-dumper.py /tmp/massive-symbol-zips
165.6MB/s            314.6MB              1.900s
163.3MB/s            287.7MB              1.762s
155.2MB/s            278.6MB              1.795s
140.9MB/s            513.2MB              3.643s
137.4MB/s            556.9MB              4.052s
134.6MB/s            531.0MB              3.946s
165.7MB/s            314.2MB              1.897s
158.1MB/s            301.5MB              1.907s
151.6MB/s            253.8MB              1.674s
146.9MB/s            502.7MB              3.422s
163.7MB/s            288.0MB              1.759s


Average speed: 153.0MB/s
Median speed:  155.2MB/s

Average files created:       150
Average directories created: 119

So that is, on average, 10% faster than c4.4xlarge.

Is it 10% more expensive? For a 1-year reserved instance, it's $0.796 versus $0.68 respectively. I.e. 15% more expensive. In other words, in this context it's 15% more $$$ for 10% more processing power.

UPDATE (Jan 24, 2018)

I can almost not believe it!

Thanks you Oliver who discovered (see comment below) a blaring mistake in my last conclusion. The for reserved instances (which is what we use on my Mozilla production servers) the c5.4xlarge is actually cheaper than c4.4xlarge. What?!

In my previous update I compared c4.4xlarge and c5.4xlarge and concluded that c5.4xlarge is 10% faster but 15% more expensive. That actually made sense. Fancier servers, more $$$. But it's not like that in the real world. See for yourself:

c4.4xlarge

c5.4xlarge

Last week I released django-memoize-function which is a library for Django developers to more conveniently use caching in function calls. This is a quick blog post to demonstrate that with an example.

The verbose traditional way to do it

Suppose you have a view function that takes in a request and returns a HttpResponse. Within, it does some expensive calculation that you know could be cached. Something like this:

No caching

def blog_post(request, slug):
    post = BlogPost.objects.get(slug=slug)

    related_posts = BlogPost.objects.exclude(
        id=post.id
    ).filter(
        # BlogPost.keywords is an ArrayField
        keywords__overlap=post.keywords
    ).order_by('-publish_date')

    context = {
        'post': post,
        'related_posts': related_posts,
    }
    return render(request, 'blogpost.html', context)

So far so good. Perhaps you know that lookup of related posts is slowish and can be cached for at least one hour. So you add this:

Caching

from django.core.cache import cache

def blog_post(request, slug):
    post = BlogPost.objects.get(slug=slug)

    cache_key = b'related_posts:{}'.format(post.id)
    related_posts = cache.get(cache_key)
    if related_posts is None:  # was not cached
        related_posts = BlogPost.objects.exclude(
            id=post.id
        ).filter(
            # BlogPost.keywords is an ArrayField
            keywords__overlap=post.keywords
        ).order_by('-publish_date')
        cache.set(cache_key, related_posts, 60 * 60)

    context = {
        'post': post,
        'related_posts': related_posts,
    }
    return render(request, 'blogpost.html', context)

Great progress. But now you want that cache to immediate reset as soon as the blog posts change.

@login_required
def update_blog_post(request, slug):
    post = BlogPost.objects.get(slug=slug)
    if request.method == 'POST':
        # BlogPostForm is a forms.ModelForm class for BlogPost
        form = BlogPostForm(request.POST, instance=post)
        if form.is_valid():
            form.save()
            cache_key = b'related_posts:{}'.format(post.id)
            cache.delete(cache_key)
            return redirect(reverse('blog_post', args=(post.id,))
    else:
        form = BlogPostForm(instance=post)

    context = {
        'post': post,
        'form': form,
    }
    return render(request, 'edit_blogpost.html', context)

Awesome. Now the cache is cleared as soon as the BlogPost is updated.

Problem; you have repeated the code generating the cache key in two places.

Use django-cache-memoize

First extract out the getting of related posts into its own function and then decorate it.

from cache_memoize import cache_memoize

@cache_memoize(60 * 60)
def get_related_posts(id):
    return BlogPost.objects.exclude(
        id=post.id
    ).filter(
        # BlogPost.keywords is an ArrayField
        keywords__overlap=post.keywords
    ).order_by('-publish_date')


def blog_post(request, slug):
    post = BlogPost.objects.get(slug=slug)

    related_posts = get_related_posts(post.id)

    context = {
        'post': post,
        'related_posts': related_posts,
    }
    return render(request, 'blogpost.html', context) 

Now, to do the cache invalidation you need to call that function get_related_posts one more time:

def update_blog_post(request, slug):
    post = BlogPost.objects.get(slug=slug)
    if request.method == 'POST':
        # BlogPostForm is a forms.ModelForm class for BlogPost
        form = BlogPostForm(request.POST, instance=post)
        if form.is_valid():
            form.save()

            # NOTE!
            get_related_posts.invalidate(post.id)

            return redirect(reverse('blog_post', args=(post.id,))
    else:
        form = BlogPostForm(instance=post)

    context = {
        'post': post,
        'form': form,
    }
    return render(request, 'edit_blogpost.html', context)

Now you're not repeating the code that constructs the cache key.

Getting fancy; hot cache

The above pattern, with or without django-cache-memoize, clears the cache when the blog post changes and then you basically wait till the next time the blog post is rendered, then the cache will be populated again.

A more "aggressive" pattern is to "heat the cache up" right after we've cleared it. A simple change is to call get_related_posts() again and let it cache. But to make sure it gets a fresh set of results we pass in the extra _refresh=True argument.

def update_blog_post(request, slug):
    post = BlogPost.objects.get(slug=slug)
    if request.method == 'POST':
        # BlogPostForm is a forms.ModelForm class for BlogPost
        form = BlogPostForm(request.POST, instance=post)
        if form.is_valid():
            form.save()

            # NOTE!
            # Refresh the cache here and now
            get_related_posts(post.id, _refresh=True)

            return redirect(reverse('blog_post', args=(post.id,))
    else:
        form = BlogPostForm(instance=post)

    context = {
        'post': post,
        'form': form,
    }
    return render(request, 'edit_blogpost.html', context)

What was the point of that?

The above example doesn't do a great job demonstrating how convenient it can be to use django-cache-memoize compared to "doing it manually". If your code base is peppered with lots of little blocks where you construct a cache key, check the cache, fall back on re-generation and write to cache again; then it can really add up to take away all of that mess and just use a decorator on anything that can be memoized.

Probably the biggest benefit with moving the cacheable functionality into its own function and decorating it is that all that hassle code with creating safe and unique cache keys is all in one place. You won't be violating the Don't Repeat Yourself principle. This becomes especially important once the cache keys that need to be constructed are getting complex and needs care.

Ultimately if you're able, your code will be free of various cache.set and cache.get code and yet a bunch of cacheable stuff gets cached nicely.

Why not use a regular @memoize or @functools.lru_cache?

The major difference between something like https://pypi.python.org/pypi/memoize/ and django-cache-memoize is that django-cache-memoize uses django.core.cache.cache which is a global state store (most likely backed by Redis or Memcached). If you use one of the other memoization solutions they'll be in-memory. Meaning, if your production code runs Gunicorn or uWSGI with, say, 8 workers then you'll have 8 copies of the same cache store. So if you're trying to protect an expensive function with @functools.lru_cache it will, worst case, be a cache miss 8 times on 8 different requests.

Released a new package today: django-cache-memoize

• Docs
• On GitHub
• On PyPI

It's actually quite simple; a Python memoize function that uses Django's cache plus the added trick that you can invalidate the cache my doing the same function call with the same parameters if you just add .invalidate to your function.

The history of it is from my recent Mozilla work on Symbols.

I originally copy and pasted the snippet out that in a blog post and today I extracted it out into its own project with tests, docs, CI and a setup.py.

I'm still amazed how long it takes to make a package with all the "fluff" around it. A lot of the bits in here (like setup.py and pytest.ini etc) are copied from other nicely maintained Python packages. For example, I straight up copied the tox.ini from Jannis Leidel's python-dockerflow. The ratio of actual code writing (including tests!) is far overpowered by the package sit-ups. But I "complain with a pinch of salt" because a lot of time spent was writing documentation and that's equally as important as the code probably.

Suppose you have a bunch of files you need to Gzip in Python; what's the optimal way to do that? In serial, to avoid saturating the GIL? In multiprocessing, to spread the load across CPU cores? Or with threads?

I needed to know this for symbols.mozilla.org since it does a lot of Gzip'ing. In symbols.mozilla.org clients upload a zip file full of files. A lot of them are plain text and when uploaded to S3 it's best to store them gzipped. Basically it does this:

def upload_sym_file(s3_client, payload, bucket_name, key_name):
    file_buffer = BytesIO()
    with gzip.GzipFile(fileobj=file_buffer, mode='w') as f:
        f.write(payload)
    file_buffer.seek(0, os.SEEK_END)
    size = file_buffer.tell()
    file_buffer.seek(0)
    s3_client.put_object(
        Bucket=bucket_name,
        Key=key_name,
        Body=file_buffer
    )
    print(f"Uploaded {size}")

Another important thing to consider before jumping into the benchmark is to appreciate the context of this application; the bundles of files I need to gzip are often many but smallish. The average file size of the files that need to be gzip'ed is ~300KB. And each bundle is between 5 to 25 files.

The Benchmark

For the sake of the benchmark, here, all it does it figure out the size of each gzipped buffer and reports that as a list.

f1 - Basic serial

def f1(payloads):
    sizes = []
    for payload in payloads:
        sizes.append(_get_size(payload))
    return sizes

f2 - Using multiprocessing.Pool

def f2(payloads):  # multiprocessing
    sizes = []
    with multiprocessing.Pool() as p:
        sizes = p.map(_get_size, payloads)
    return sizes

f3 - Using concurrent.futures.ThreadPoolExecutor

def f3(payloads):  # concurrent.futures.ThreadPoolExecutor
    sizes = []
    futures = []
    with concurrent.futures.ThreadPoolExecutor() as executor:
        for payload in payloads:
            futures.append(
                executor.submit(
                    _get_size,
                    payload
                )
            )
        for future in concurrent.futures.as_completed(futures):
            sizes.append(future.result())
    return sizes

f4 - Using concurrent.futures.ProcessPoolExecutor

def f4(payloads):  # concurrent.futures.ProcessPoolExecutor
    sizes = []
    futures = []
    with concurrent.futures.ProcessPoolExecutor() as executor:
        for payload in payloads:
            futures.append(
                executor.submit(
                    _get_size,
                    payload
                )
            )
        for future in concurrent.futures.as_completed(futures):
            sizes.append(future.result())
    return sizes

Note that when using asynchronous methods like this, the order of items returned is not the same as they're submitted. An easy remedy if you need the results back in order is to not use a list but to use a dictionary. Then you can track each key (or index if you like) to a value.

The Results

I ran this on three different .zip files of different sizes. To get some sanity in the benchmark I made it print out how many bytes it has to process and how many bytes the gzip will manage to do.

# files 66
Total bytes to gzip 140.69MB
Total bytes gzipped 14.96MB
Total bytes shaved off by gzip 125.73MB

# files 103
Total bytes to gzip 331.57MB
Total bytes gzipped 66.90MB
Total bytes shaved off by gzip 264.67MB

# files 26
Total bytes to gzip 86.91MB
Total bytes gzipped 8.28MB
Total bytes shaved off by gzip 78.63MB

Sorry for being eastetically handicapped when it comes to using Google Docs but here goes...

This demonstrates the median times it takes each function to complete, each of the three different files.

In all three files I tested, clearly doing it serially (f1) is the worst. Supposedly since my laptop has more than one CPU core and the others are not being used. Another pertinent thing to notice is that when the work is really big, (the middle 4 bars) the difference isn't as big doing things serially compared to concurrently.

That second zip file contained a single file that was 80MB. The largest in the other two files were 18MB and 22MB.

This is the mean across all medians grouped by function and each compared to the slowest.

I call this the "bestest graph". It's a combination across all different sizes and basically concludes which one is the best, which clearly is function f3 (the one using concurrent.futures.ThreadPoolExecutor).

CPU Usage

This is probably the best way to explain how the CPU is used; I ran each function repeatedly, then opened gtop and took a screenshot of the list of processes sorted by CPU percentage.

f1 - Serially

No distractions but it takes 100% of one CPU to work.

f2 - multiprocessing.Pool

My laptop has 8 CPU cores, but I don't know why I see 9 Python processes here.
I don't know why each CPU isn't 100% but I guess there's some administrative overhead to start processes by Python.

f3 - concurrent.futures.ThreadPoolExecutor

One process, with roughly 5 x 8 = 40 threads GIL swapping back and forth but all in all it manages to keep itself very busy since threads are lightweight to share data to.

f4 - concurrent.futures.ProcessPoolExecutor

This is actually kinda like multiprocessing.Pool but with a different (arguably easier) API.

Conclusion

By a small margin concurrent.futures.ThreadPoolExecutor won. That's despite not being able to use all CPU cores. This, pseudo scientifically, proves that the overhead of starting the threads is (remember average number of files in each .zip is ~65) more worth it than being able to use all CPUs.

Discussion

There's an interesting twist to this! At least for my use case...

In the application I'm working on, there's actually a lot more that needs to be done other than just gzip'ping some blobs of files. For each file I need to a HEAD query to AWS S3 and an PUT query to AWS S3 too. So what I actually need to do is create an instance of client = botocore.client.S3 that I use to call client.list_objects_v2 and client.put_object.

When you create an instance of botocore.client.S3, automatically botocore will instanciate itself with credentials from os.environ['AWS_ACCESS_KEY_ID'] etc. (or read from some /.aws file). Once created, if you ask it to do many different network operations, internally it relies on urllib3.poolmanager.PoolManager which is a list of 10 HTTP connections that get reused.

So when you run the serial version you can re-use the client instance for every file you process but you can only use one HTTP connection in the pool. With the concurrent.futures.ThreadPoolExecutor it can not only re-use the same instance of botocore.client.S3 it can cycle through all the HTTP connections in the pool.

The process based alternatives like multiprocessing.Pool and concurrent.futures.ProcessPoolExecutor can not re-use the botocore.client.S3 instance since it's not pickle'able. And it has to create a new HTTP connection for every single file.

So, the conclusion of the above rambling is that concurrent.futures.ThreadPoolExecutor is really awesome! Not only did it perform excellently in the Gzip benchmark, it has the added bonus that it can share instance objects and HTTP connections.

As of PostgreSQL 9.5 we have UPSERT support. Technically, it's ON CONFLICT, but it's basically a way to execute an UPDATE statement in case the INSERT triggers a conflict on some column value. By the way, here's a great blog post that demonstrates how to use ON CONFLICT.

In this Django app I have a model that has a field called hash which has a unique=True index on it. What I want to do is either insert a row, or if the hash is already in there, it should increment the count and the modified_at timestamp instead.

The Code(s)

Here's the basic version in "pure Django ORM":

if MissingSymbol.objects.filter(hash=hash_).exists():
    MissingSymbol.objects.filter(hash=hash_).update(
        count=F('count') + 1,
        modified_at=timezone.now()
    )
else:
    MissingSymbol.objects.create(
        hash=hash_,
        symbol=symbol,
        debugid=debugid,
        filename=filename,
        code_file=code_file or None,
        code_id=code_id or None,
    )

Here's that same code rewritten in "pure SQL":

from django.db import connection


with connection.cursor() as cursor:
    cursor.execute("""
        INSERT INTO download_missingsymbol (
            hash, symbol, debugid, filename, code_file, code_id,
            count, created_at, modified_at
        ) VALUES (
            %s, %s, %s, %s, %s, %s,
            1, CLOCK_TIMESTAMP(), CLOCK_TIMESTAMP()
          )
        ON CONFLICT (hash)
        DO UPDATE SET
            count = download_missingsymbol.count + 1,
            modified_at = CLOCK_TIMESTAMP()
        WHERE download_missingsymbol.hash = %s
        """, [
            hash_, symbol, debugid, filename,
            code_file or None, code_id or None,
            hash_
        ]
    )

Both work.

Note the use of CLOCK_TIMESTAMP() instead of NOW(). Since Django wraps all writes in transactions if you use NOW() it will be evaluated to the same value for the whole transaction, thus making unit testing really hard.

But which is fastest?

The Results

First of all, this hard to test locally because my Postgres is running locally in Docker so the network latency in talking to a network Postgres means that the latency is less and having to do two different executions would cost more if the network latency is more.

I ran a simple benchmark where it randomly picked one of the two code blocks (above) depending on a 50% chance.
The results are:

METHOD     MEAN       MEDIAN
SQL        6.99ms     6.61ms
ORM        10.28ms    9.86ms

So doing it with a block of raw SQL instead is 1.5 times faster. But this difference would surely grow when the network latency is higher.

Discussion

There's an alternative and that's to use django-postgres-extra but I'm personally hesitant. The above little raw SQL hack is the only thing I need and adding more libraries makes far-future maintenance harder.

Beyond the time optimization of being able to send only 1 SQL instruction to PostgreSQL, the biggest benefit is avoiding concurrency race conditions. From the documentation:

"ON CONFLICT DO UPDATE guarantees an atomic INSERT or UPDATE outcome; provided there is no independent error, one of those two outcomes is guaranteed, even under high concurrency. This is also known as UPSERT — "UPDATE or INSERT"."

I'm going to keep this little hack. It's not beautiful but it works and saves time and gives me more comfort around race conditions.

UPDATE - Oct 27, 2017 This snippet did now become its own PyPI package. See https://pypi.python.org/pypi/django-cache-memoize

This is something that's grown up organically when working on Mozilla Symbol Server. It has served me very well and perhaps it's worth extracting into its own lib.

Usage

Basically, you are probably used to this in Django:

from django.core.cache import cache

def compute_something(user, special=False):
    cache_key = 'meatycomputation:{}:special={}'.format(user.id, special)
    value = cache.get(cache_key)
    if value is None:
        value = _call_the_meat(user.id, special)  # some really slow function
        cache.set(cache_key, value, 60 * 5)
    return value

Here's instead how you can do exactly the same with cache_memoize:

from wherever.decorators import cache_memoize

@cache_memoize(60 * 5)
def compute_something(user, special=False):
    return _call_the_meat(user.id, special)  # some really slow function

Cache invalidation

If you ever need to do non-trivial caching you know it's important to be able to invalidate the cache. Usually, to be able to do that you need to involved in how the cache key was created.

Consider our two examples above, here's first the common thing to do:

def save_user(user):
    do_something_that_will_need_to_cache_invalidate(user)

    cache_key = 'meatycomputation:{}:special={}'.format(user.id, False)
    cache.delete(cache_key)
    # And when it was special=True
    cache_key = 'meatycomputation:{}:special={}'.format(user.id, True)
    cache.delete(cache_key)

This works but it involves repeating the code that generates the cache key. You could extract that into its own function of course.

Here's how you do it with the cache_memoize decorator:

def save_user(user):
    do_something_that_will_need_to_cache_invalidate(user)

    compute_something.invalidate(user, special=False)
    compute_something.invalidate(user, special=True)    

Other features

There are actually two ways to "invalidate" the cache. Calling the new myoriginalfunction.invalidate(...) function or passing a custom extra keyword argument called _refresh. For example: compute_something(user, _refresh=True).

You can pass callables that get called when the cache works in your favor or when it's a cache miss. For example:

def increment_hits(user, special=None):
    # use your imagination
    metrics.incr(user.email)

def cache_miss(user, special=None):
    print("cache miss on {}".format(user.email))

@cache_memoize(
    60 * 5,
    hit_callable=increment_hits,
    miss_callable=cache_miss,
)
def compute_something(user, special=False):
    return _call_the_meat(user.id, special)  # some really slow function

Sometimes you just want to use the memoizer to make sure something only gets called "once" (or once per time interval). In that case it might be smart to not flood your cache backend with the value of the function output if there is one. For example:

@cache_memoize(60 * 60, store_result=False)  # idempotent guard
def calculate_and_update(user):
    # do something expensive here that is best to only do once per hour

Internally cache_memoize will basically try to convert every argument and keyword argument to a string with, kinda, str(). That might not always be appropriate because you might know that you have two distinct objects whose __str__ will yield the same result. For that you can use the args_rewrite parameter. For example:

def simplify_special_objects(obj):
    # use your imagination
    return obj.hostname 

@cache_memoize(60 * 5, args_rewrite=simplify_special_objects)
def compute_something(special_obj):
    return _call_the_meat(special_obj.hostname)

In conclusion

I've uploaded the code as a gist.

It's quite possible that there's already a perfectly good lib that does exactly this. If so, thanks for letting me know. If not, perhaps I ought to wrap this up and publish it on PyPI. Again, that's for letting me know.

UPDATE

I found a bug in the original gist. Updated 2017-10-05.
The bug was that the calling of miss_callable and hit_callable was reversed.

(Thanks Miles Crabil not only for being an awesome Ops person but also for reviewing this blog post!)

My project over the summer, here at Mozilla, has been a project called Mozilla Symbol Server. It's a web service that uploads C++ symbol files, downloads C++ symbol files and symbolicates C++ crash stacktraces. It went into production last week which was fun but there's still lots of work to do on adding beyond-parity features and more optimizations.

What Is Mozilla Symbol Server?

The code name for this project is Tecken and it's written in Python (Django, Gunicorn) and uses PostgreSQL, Redis and Celery. The frontend is entirely static and developed (almost) as a separate project within. The frontend is written in React (using create-react-app and react-router). Everything is run as Docker containers. And if you ask me more details about how it's configured/deployed I'm afraid I have to defer to the awesome Mozilla CloudOps team.

One the challenges I faces developing Tecken is that symbol downloads need to be fast to handle high volumes of traffic. Today I did some load testing on our stage deployment and managed to start 14 concurrent clients that bombarded our staging server with realistic HTTPS GET queries based on log files. It's actually 7 + 1 + 4 + 2 concurrent clients. 7 of them from a m3.2xlarge EC2 node (8 vCPUs), 1 from a m3.large EC2 node (1 vCPU), 2 from two separate NYC based DigitalOcean personal servers and 2 clients here from my laptop on my home broadband. Basically, each loadtest script process got its own CPU.

It's hard to know how much more each client could push if it wasn't slowed down. Either way, the server managed to sustain about 330 requests per second. Our production baseline goal is to able to handle at least 40 requests per second.

After running for a while the caches started getting warm but about 1-5% of requests do have to make a boto3 roundtrip to an S3 bucket located on the other side of America in Oregon. There is also a ~5% penalty in that some requests trigger a write to a central Redis ElastiCache server. That's cheaper than the boto3 S3 call but still hefty latency costs to pay.

The ELB in our staging environment spreads the load between 2 c4.large (2 vCPUs, 3.75GB RAM) EC2 web heads. Each running with preloaded Gunicorn workers between Nginx and Django. Each web head has its own local memcached server to share memory between each worker but only local to the web head.

Is this a lot?

How long is a rope? Hard to tell. Tecken's performance is certainly more than enough and by the sheer fact that it was only just production deployed last week tells me we can probably find a lot of low-hanging fruit optimizations on the deployment side over time.

One way of answering that is to compare it with our lightest endpoint. One that involves absolutely no external resources. It's just pure Python in the form of ELB → Nginx → Gunicorn → Django. If I run hey from the same server I did the load testing I get a topline of 1,300 requests per second.

$ hey -n 10000 -c 10 https://symbols.stage.mozaws.net/__lbheartbeat__
Summary:
  Total:    7.6604 secs
  Slowest:  0.0610 secs
  Fastest:  0.0018 secs
  Average:  0.0075 secs
  Requests/sec: 1305.4199
...

That basically means that all the extra "stuff" (memcache key prep, memcache key queries and possible other high latency network requests) it needs to do in the Django view takes up roughly 3x the time it takes the absolute minimal Django request-response rendering.

Also, if I use the same technique to bombard a single URL, but one that actually involves most code steps but is definitely able to not require any slow ElastiCache writes or boto3 S3 reads you I get 800 requests per second:

$ hey -n 10000 -c 10 https://symbols.stage.mozaws.net/advapi32.pdb/5EFB9BF42CC64024AB64802E467394642/advapi32.sy
Summary:
  Total:    12.4160 secs
  Slowest:  0.0651 secs
  Fastest:  0.0024 secs
  Average:  0.0122 secs
  Requests/sec: 805.4150
  Total data:   300000 bytes
  Size/request: 30 bytes
...

Lesson learned

It's a recurring reminder that performance is almost all about latency. If not RAM or disk it's networking. See the graph of the "Max CPU Used" which basically shows that CPU of user, system and stolen ("CPU spent waiting for the hypervisor to service another virtual CPU") never sum totalling over 50%.

tl;dr; By a slim margin, the fastest way to check a filename matching a list of extensions is filename.endswith(extensions)

This turned out to be premature optimization. The context is that I want to check if a filename matches the file extension in a list of 6.

The list being ['.sym', '.dl_', '.ex_', '.pd_', '.dbg.gz', '.tar.bz2']. Meaning, it should return True for foo.sym or foo.dbg.gz. But it should return False for bar.exe or bar.gz.

I put together a litte benchmark, ran it a bunch of times and looked at the results. Here are the functions I wrote:

def f1(filename):
    for each in extensions:
        if filename.endswith(each):
            return True
    return False


def f2(filename):
    return filename.endswith(extensions_tuple)


regex = re.compile(r'({})$'.format(
    '|'.join(re.escape(x) for x in extensions)
))


def f3(filename):
    return bool(regex.findall(filename))


def f4(filename):
    return bool(regex.search(filename))

The results are boring. But I guess that's a result too:

FUNCTION             MEDIAN               MEAN
f1 9543 times        0.0110ms             0.0116ms
f2 9523 times        0.0031ms             0.0034ms
f3 9560 times        0.0041ms             0.0045ms
f4 9509 times        0.0041ms             0.0043ms

For a list of ~40,000 realistic filenames (with result True 75% of the time), I ran each function 10 times. So, it means it took on average 0.0116ms to run f1 10 times here on my laptop with Python 3.6.

More premature optimization

Upon looking into the data and thinking about this will be used. If I reorder the list of extensions so the most common one is first, second most common second etc. Then the performance improves a bit for f1 but slows down slightly for f3 and f4.

Conclusion

That .endswith(some_tuple) is neat and it's hair-splittingly faster. But really, this turned out to not make a huge difference in the grand scheme of things. On average it takes less than 0.001ms to do one filename match.

I did another couple of benchmarks of different cache backends in Django. This is an extension/update on Fastest cache backend possible for Django published a couple of months ago. This benchmarking isn't as elaborate as the last one. Fewer tests and fewer variables.

I have another app where I use a lot of caching. This web application will run its cache server on the same virtual machine. So no separation of cache server and web head(s). Just one Django server talking to localhost:11211 (memcached's default port) and localhost:6379 (Redis's default port).

Also in this benchmark, the keys were slightly smaller. To simulate my applications "realistic needs" I made the benchmark fall on roughly 80% cache hits and 20% cache misses. The cache keys were 1 to 3 characters long and the cache values lists of strings always 30 items long (e.g. len(['abc', 'def', 'cba', ... , 'cab']) == 30).

Also, in this benchmark I was too lazy to test all different parsers, serializers and compressors that django-redis supports. I only test python-memcached==1.58 versus django-redis==4.8.0 versus django-redis==4.8.0 && msgpack-python==0.4.8.

The results are quite "boring". There's basically not enough difference to matter.

Config	Average	Median	Compared to fastest
memcache	4.51s	3.90s	100%
redis	5.41s	4.61s	84.7%
redis_msgpack	5.16s	4.40s	88.8%

UPDATE

As Hal pointed out in the comment, when you know the web server and the memcached server is on the same computer you should use UNIX sockets. They're "obviously" faster since the lack of HTTP overhead at the cost of it doesn't work over a network.

Because running memcached on a socket on OSX is a hassle I only have one benchmark. Note! This basically compares good old django.core.cache.backends.memcached.MemcachedCache with two different locations.

Config	Average	Median	Compared to fastest
127.0.0.1:11211	3.33s	3.34s	81.3%
unix:/tmp/memcached.sock	2.66s	2.71s	100%

But there's more! Another option is to use pylibmc which is a Python client written in C. By the way, my Python I use for these microbenchmarks is Python 3.5.

Unfortunately I'm too lazy/too busy to do a matrix comparison of pylibmc on TCP versus UNIX socket. Here are the comparison results of using python-memcached versus pylibmc:

Client	Average	Median	Compared to fastest
python-memcached	3.52s	3.52s	62.9%
pylibmc	2.31s	2.22s	100%

UPDATE 2

Seems my luck someone else has done the matrix comparison of python-memcached vs pylibmc on TCP vs UNIX socket:

https://plot.ly/~jensens/36.embed

If you're reading this you're probably familiar with how, in django-pipeline, you define bundles of static files to be combined and served. If you're not familiar with django-pipeline it's unlike this'll be of much help.

The Challenge (aka. the pitfall)

So you specify bundles by creating things in your settings.py something like this:

PIPELINE = {
    'STYLESHEETS': {
        'colors': {
            'source_filenames': (
              'css/core.css',
              'css/colors/*.css',
              'css/layers.css'
            ),
            'output_filename': 'css/colors.css',
            'extra_context': {
                'media': 'screen,projection',
            },
        },
    },
    'JAVASCRIPT': {
        'stats': {
            'source_filenames': (
              'js/jquery.js',
              'js/d3.js',
              'js/collections/*.js',
              'js/aplication.js',
            ),
            'output_filename': 'js/stats.js',
        }
    }
}

You do a bit more configuration and now, when you run ./manage.py collectstatic --noinput Django and django-pipeline will gather up all static files from all Django apps installed, then start post processing then and doing things like concatenating them into one file and doing stuff like minification etc.

The problem is, if you look at the example snippet above, there's a typo. Instead of js/application.js it's accidentally js/aplication.js. Oh noes!!

What's sad is it that nobody will notice (running ./manage.py collectstatic will exit with a 0). At least not unless you do some careful manual reviewing. Perhaps you will notice later, when you've pushed the site to prod, that the output file js/stats.js actually doesn't contain the code from js/application.js.

Or, you can automate it!

A Solution (aka. the hack)

I started this work this morning because the error actually happened to us. Thankfully not in production but our staging server produced a rendered HTML page with <link href="/static/css/report.min.cd784b4a5e2d.css" rel="stylesheet" type="text/css" /> which was an actual file but it was 0 bytes.

It wasn't that hard to figure out what the problem was because of the context of recent changes but it would have been nice to catch this during continuous integration.

So what we did was add an extra class to settings.STATICFILES_FINDERS called myproject.base.finders.LeftoverPipelineFinder. So now it looks like this:

# in settings.py

STATICFILES_FINDERS = (
    'django.contrib.staticfiles.finders.FileSystemFinder',
    'django.contrib.staticfiles.finders.AppDirectoriesFinder',
    'pipeline.finders.PipelineFinder',
    'myproject.finders.LeftoverPipelineFinder',  # the new hotness!
)

And here's the class implementation:

from pipeline.finders import PipelineFinder

from django.conf import settings
from django.core.exceptions import ImproperlyConfigured


class LeftoverPipelineFinder(PipelineFinder):
    """This finder is expected to come AFTER 
    django.contrib.staticfiles.finders.FileSystemFinder and 
    django.contrib.staticfiles.finders.AppDirectoriesFinder in 
    settings.STATICFILES_FINDERS.
    If a path is looked for here it means it's trying to find a file
    that none of the regular staticfiles finders couldn't find.
    """
    def find(self, path, all=False):
        # Before we raise an error, try to find out where,
        # in the bundles, this was defined. This will make it easier to correct
        # the mistake.
        for config_name in 'STYLESHEETS', 'JAVASCRIPT':
            config = settings.PIPELINE[config_name]
            for key in config:
                if path in config[key]['source_filenames']:
                    raise ImproperlyConfigured(
                        'Static file {!r} can not be found anywhere. Defined in '
                        "PIPELINE[{!r}][{!r}]['source_filenames']".format(
                            path,
                            config_name,
                            key,
                        )
                    )
        # If the file can't be found AND it's not in bundles, there's
        # got to be something else really wrong.
        raise NotImplementedError(path)

Now, if you have a typo or something in your bundles, you'll get a nice error about it as soon as you try to run collectstatic. For example:

▶ ./manage.py collectstatic --noinput
Post-processed 'css/search.min.css' as 'css/search.min.css'
Post-processed 'css/base.min.css' as 'css/base.min.css'
Post-processed 'css/base-dynamic.min.css' as 'css/base-dynamic.min.css'
Post-processed 'js/google-analytics.min.js' as 'js/google-analytics.min.js'
Traceback (most recent call last):
...
django.core.exceptions.ImproperlyConfigured: Static file 'js/aplication.js' can not be found anywhere. Defined in PIPELINE['JAVASCRIPT']['stats']['source_filenames']

Final Thoughts

This was a morning hack. I'm still not entirely sure if this the best approach, but there was none better and the result is pretty good.

We run ./manage.py collectstatic --noinput in our continous integration just before it runs ./manage.py test. So if you make a Pull Request that has a typo in bundles.py it will get caught.

Unfortunately, it won't find missing files if you use foo*.js or something like that. django-pipeline uses glob.glob to convert expressions like that into a list of actual files and that depends on the filesystem and all of that happens before the django.contrib.staticfiles.finders.find function is called.

If you have any better suggestions to solve this, please let me know.