Peterbe.com

Last week, I blogged about "How much faster is Redis at storing a blob of JSON compared to PostgreSQL?". Judging from a lot of comments, people misinterpreted this. (By the way, Redis is persistent). It's no surprise that Redis is faster.

However, it's a fact that I have do have a lot of blobs stored and need to present them via the web API as fast as possible. It's rare that I want to do relational or batch operations on the data. But Redis isn't a slam dunk for simple retrieval because I don't know if I trust its integrity with the 3GB worth of data that I both don't want to lose and don't want to load all into RAM.

But is it entirely wrong to look at WHICH database to get the best speed?

Reviewing this corner of Song Search helped me rethink this. PostgreSQL is, in my view, a better database for storing stuff. Redis is faster for individual lookups. But you know what's even faster? Nginx

Nginx??

The way the application works is that a React web app is requesting the Amazon product data for the sake of presenting an appropriate affiliate link. This is done by the browser essentially doing:

const response = await fetch('https://songsear.ch/api/song/5246889/amazon');

Internally, in the app, what it does is that it looks this up, by ID, on the AmazonAffiliateLookup ORM model. Suppose it wasn't there in the PostgreSQL, it uses the Amazon Affiliate Product Details API, to look it up and when the results come in it stores a copy of this in PostgreSQL so we can re-use this URL without hitting rate limits on the Product Details API. Lastly, in a piece of Django view code, it carefully scrubs and repackages this result so that only the fields used by the React rendering code is shipped between the server and the browser. That "scrubbed" piece of data is actually much smaller. Partly because it limits the results to the first/best match and it deletes a bunch of things that are never needed such as ProductTypeName, Studio, TrackSequence etc. The proportion is roughly 23x. I.e. of the 3GB of JSON blobs stored in PostgreSQL only 130MB is ever transported from the server to the users.

Again, Nginx?

Nginx has a built in reverse HTTP proxy cache which is easy to set up but a bit hard to do purges on. The biggest flaw, in my view, is that it's hard to get a handle of how much RAM this it's eating up. Well, if the total possible amount of data within the server is 130MB, then that is something I'm perfectly comfortable to let Nginx handle cache in RAM.

Good HTTP performance benchmarking is hard to do but here's a teaser from my local laptop version of Nginx:

▶ hey -n 10000 -c 10 https://songsearch.local/api/song/1810960/affiliate/amazon-itunes

Summary:
  Total:    0.9882 secs
  Slowest:  0.0279 secs
  Fastest:  0.0001 secs
  Average:  0.0010 secs
  Requests/sec: 10119.8265


Response time histogram:
  0.000 [1] |
  0.003 [9752]  |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.006 [108]   |
  0.008 [70]    |
  0.011 [32]    |
  0.014 [8] |
  0.017 [12]    |
  0.020 [11]    |
  0.022 [1] |
  0.025 [4] |
  0.028 [1] |


Latency distribution:
  10% in 0.0003 secs
  25% in 0.0006 secs
  50% in 0.0008 secs
  75% in 0.0010 secs
  90% in 0.0013 secs
  95% in 0.0016 secs
  99% in 0.0068 secs

Details (average, fastest, slowest):
  DNS+dialup:   0.0000 secs, 0.0001 secs, 0.0279 secs
  DNS-lookup:   0.0000 secs, 0.0000 secs, 0.0026 secs
  req write:    0.0000 secs, 0.0000 secs, 0.0011 secs
  resp wait:    0.0008 secs, 0.0001 secs, 0.0206 secs
  resp read:    0.0001 secs, 0.0000 secs, 0.0013 secs

Status code distribution:
  [200] 10000 responses

10,000 requests across 10 clients at rougly 10,000 requests per second. That includes doing all the HTTP parsing, WSGI stuff, forming of a SQL or Redis query, the deserialization, the Django JSON HTTP response serialization etc. The cache TTL is controlled by simply setting a Cache-Control HTTP header with something like max-age=86400.

Now, repeated fetches for this are cached at the Nginx level and it means it doesn't even matter how slow/fast the database is. As long as it's not taking seconds, with a long Cache-Control, Nginx can hold on to this in RAM for days or until the whole server is restarted (which is rare).

Conclusion

If you the total amount of data that can and will be cached is controlled, putting it in a HTTP reverse proxy cache is probably order of magnitude faster than messing with chosing which database to use.

tl;dr; Redis is 16 times faster at reading these JSON blobs.

In Song Search when you've found a song, it loads some affiliate links to Amazon.com. (In case you're curious it's earning me lower double-digit dollars per month). To avoid overloading the Amazon Affiliate Product API, after I've queried their API, I store that result in my own database along with some metadata. Then, the next time someone views that song page, it can read from my local database. With me so far?

The other caveat is that you can't store these lookups locally too long since prices change and/or results change. So if my own stored result is older than a couple of hundred days, I delete it and fetch from the network again. My current implementation uses PostgreSQL (via the Django ORM) to store this stuff. The model looks like this:

class AmazonAffiliateLookup(models.Model, TotalCountMixin):
    song = models.ForeignKey(Song, on_delete=models.CASCADE)
    matches = JSONField(null=True)
    search_index = models.CharField(max_length=100, null=True)
    lookup_seconds = models.FloatField(null=True)
    created = models.DateTimeField(auto_now_add=True, db_index=True)
    modified = models.DateTimeField(auto_now=True)

At the moment this database table is 3GB on disk.

Then, I thought, why not use Redis for this. Then I can use Redis's "natural" expiration by simply setting as expiry time when I store it and then I don't have to worry about cleaning up old stuff at all.

The way I'm using Redis in this project is as a/the cache backend and I have it configured like this:

CACHES = {
    "default": {
        "BACKEND": "django_redis.cache.RedisCache",
        "LOCATION": REDIS_URL,
        "TIMEOUT": config("CACHE_TIMEOUT", 500),
        "KEY_PREFIX": config("CACHE_KEY_PREFIX", ""),
        "OPTIONS": {
            "COMPRESSOR": "django_redis.compressors.zlib.ZlibCompressor",
            "SERIALIZER": "django_redis.serializers.msgpack.MSGPackSerializer",
        },
    }
}

The speed difference

Perhaps unrealistic but I'm doing all this testing here on my MacBook Pro. The connection to Postgres (version 11.4) and Redis (3.2.1) are both on localhost.

Reads

The reads are the most important because hopefully, they happen 10x more than writes as several people can benefit from previous saves.

I changed my code so that it would do a read from both databases and if it was found in both, write down their time in a log file which I'll later summarize. Results are as follows:

PG:
median: 8.66ms
mean  : 11.18ms
stdev : 19.48ms

Redis:
median: 0.53ms
mean  : 0.84ms
stdev : 2.26ms

(310 measurements)

It means, when focussing on the median, Redis is 16 times faster than PostgreSQL at reading these JSON blobs.

Writes

The writes are less important but due to the synchronous nature of my Django, the unlucky user who triggers a look up that I didn't have, will have to wait for the write before the XHR request can be completed. However, when this happens, the remote network call to the Amazon Product API is bound to be much slower. Results are as follows:

PG:
median: 8.59ms
mean  : 8.58ms
stdev : 6.78ms

Redis:
median: 0.44ms
mean  : 0.49ms
stdev : 0.27ms

(137 measurements)

It means, when focussing on the median, Redis is 20 times faster than PostgreSQL at writing these JSON blobs.

Conclusion and discussion

First of all, I'm still a PostgreSQL fan-boy and have no intention of ceasing that. These times are made up of much more than just the individual databases. For example, the PostgreSQL speeds depend on the Django ORM code that makes the SQL and sends the query and then turns it into the model instance. I don't know what the proportions are between that and the actual bytes-from-PG's-disk times. But I'm not sure I care either. The tooling around the database is inevitable mostly and it's what matters to users.

Both Redis and PostgreSQL are persistent and survive server restarts and crashes etc. And you get so many more "batch related" features with PostgreSQL if you need them, such as being able to get a list of the last 10 rows added for some post-processing batch job.

I'm currently using Django's cache framework, with Redis as its backend, and it's a cache framework. It's not meant to be a persistent database. I like the idea that if I really have to I can just flush the cache and although detrimental to performance (temporarily) it shouldn't be a disaster. So I think what I'll do is store these JSON blobs in both databases. Yes, it means roughly 6GB of SSD storage but it also potentially means loading a LOT more into RAM on my limited server. That extra RAM usage pretty much sums of this whole blog post; of course it's faster if you can rely on RAM instead of disk. Now I just need to figure out how RAM I can afford myself for this piece and whether it's worth it.

UPDATE September 29, 2019

I experimented with an optimization of NOT turning the Django ORM query into a model instance for each record. Instead, I did this:

+from dataclasses import dataclass


+@dataclass
+class _Lookup:
+    modified: datetime.datetime
+    matches: list

...

+base_qs = base_qs.values_list("modified", "matches")
-lookup = base_qs.get(song__id=song_id)
+lookup_tuple = base_qs.get(song__id=song_id)
+lookup = _Lookup(*lookup_tuple)

print(lookup.modified)

Basically, let the SQL driver's "raw Python" content come through the Django ORM. The old difference between PostgreSQL and Redis was 16x. The new difference was 14x instead.

By analyzing my Nginx logs, I've concluded that SongSearch's autocomplete JSON API now gets about 2.2 requests per second. I.e. these are XHR requests to /api/search/autocomplete?q=....

Roughly, 1.8 requests per second goes back to the Django/Elasticsearch backend. That's a hit ratio of 16%. These Django/Elasticsearch requests take roughly 200ms on average. I suspect about 150-180ms of that time is spent querying Elasticsearch, the rest being Python request/response and JSON "paperwork".

Caching strategy

Caching is hard because the queries are so vastly different over time. Had I put a Redis cache decorator on the autocomplete Django view function I'd quickly bloat Redis memory and cause lots of evictions.

What I used to do was something like this:

def search_autocomplete(request):
   q = request.GET.get('q') 

   cache_key = None
   if len(q) < 10:
      cache_key = 'autocomplete:' + q
      results = cache.get(cache_key)
      if results is not None:
          return http.JsonResponse(results)

   results = _do_elastisearch_query(q)
   if cache_key:
       cache.set(cache_key, results, 60 * 60)

   return http.JsonResponse(results)   

However, after some simple benchmarking it was clear that using Nginx' uwsgi_cache it was much faster to let the cacheable queries terminate already at Nginx. So I changed the code to something like this:

def search_autocomplete(request):
   q = request.GET.get('q') 
   results = _do_elastisearch_query(q)
   response = http.JsonResponse(results)   

   if len(q) < 10:
       patch_cache_control(response, public=True, max_age=60 * 60)

   return response

The only annoying thing about Nginx caching is that purging is hard unless you go for that Nginx Plus (or whatever their enterprise version is called). But more annoying, to me, is that fact that I can't really see what this means for my server. When I was caching with Redis I could just use redis-cli and...

> INFO
...
# Memory
used_memory:123904288
used_memory_human:118.16M
...

Nginx Amplify

My current best tool for keeping an eye on Nginx is Nginx Amplify. It gives me some basic insights about the state of things. Here are some recent screenshots:

Thoughts and conclusion

Caching is hard. But it's also fun because it ties directly into performance work.

In my business logic, I chose that autocomplete queries that are between 1 and 9 characters are cacheable. And I picked a TTL of 60 minutes. At this point, I'm not sure exactly why I chose that logic but I remember doing some back-of-envelope calculations about what the hit ratio would be and roughly what that would mean in bytes in RAM. I definitely remember picking 60 minutes because I was nervous about bloating Nginx's memory usage. But as of today, I'm switching that up to 24 hours and let's see what that does to my current 16% Nginx cache hit ratio. At the moment, /var/cache/nginx-cache/ is only 34MB which isn't much.

Another crux with using uwsgi_cache (or proxy_cache) is that you can't control the cache key very well. When it was all in Python I was able to decide about the cache key myself. A plausible implementation is cache_key = q.lower().strip() for example. That means you can protect your Elasticsearch backend from having to do {"q": "A"} and {"q": "a"}. Who knows, perhaps there is a way to hack this in Nginx without compiling in some Lua engine.

The ideal would be some user-friendly diagnostics tool that I can point somewhere, towards Nginx, that says how much my uwsgi_cache is hurting or saving me. Autocomplete is just one of many things going on on this single DigitalOcean server. There's also a big PostgreSQL server, a node-express cluster, a bunch of uwsgi workers, Redis, lots of cron job scripts, and of course a big honking Elasticsearch 6.

UPDATE (July 12 2019)

Currently, and as mentioned above, I only set Cache-Control headers (which means Nginx snaps it up) for queries that at max 9 characters long. I wanted to appreciate and understand how ratio of all queries are longer than 9 characters so I wrote a report and its output is this:

POINT: 7
Sum show 75646 32.2%
Sum rest 159321 67.8%

POINT: 8
Sum show 83702 35.6%
Sum rest 151265 64.4%

POINT: 9
Sum show 90870 38.7%
Sum rest 144097 61.3%

POINT: 10
Sum show 98384 41.9%
Sum rest 136583 58.1%

POINT: 11
Sum show 106093 45.2%
Sum rest 128874 54.8%

POINT: 12
Sum show 113905 48.5%
Sum rest 121062 51.5%

It means that (independent of time expiry) 38.7% of queries are 9 characters or less.

Here's the code. It's quick-n-dirty but it works wonderfully:

import functools
import hashlib

from django.core.cache import cache
from django.utils.encoding import force_bytes


def lock_decorator(key_maker=None):
    """
    When you want to lock a function from more than 1 call at a time.
    """

    def decorator(func):
        @functools.wraps(func)
        def inner(*args, **kwargs):
            if key_maker:
                key = key_maker(*args, **kwargs)
            else:
                key = str(args) + str(kwargs)
            lock_key = hashlib.md5(force_bytes(key)).hexdigest()
            with cache.lock(lock_key):
                return func(*args, **kwargs)

        return inner

    return decorator

How To Use It

This has saved my bacon more than once. I use it on functions that really need to be made synchronous. For example, suppose you have a function like this:

def fetch_remote_thing(name):
    try:
        return Thing.objects.get(name=name).result
    except Thing.DoesNotExist:
        # Need to go out and fetch this
        result = some_internet_fetching(name)  # Assume this is sloooow
        Thing.objects.create(name=name, result=result)
        return result

That function is quite dangerous because if executed by two concurrent web requests for example, they will trigger
two "identical" calls to some_internet_fetching and if the database didn't have the name already, it will most likely trigger two calls to Thing.objects.create(name=name, ...) which could lead to integrity errors or if it doesn't the whole function breaks down because it assumes that there is only 1 or 0 of these Thing records.

Easy to solve, just add the lock_decorator:

@lock_decorator()
def fetch_remote_thing(name):
    try:
        return Thing.objects.get(name=name).result
    except Thing.DoesNotExist:
        # Need to go out and fetch this
        result = some_internet_fetching(name)  # Assume this is sloooow
        Thing.objects.create(name=name, result=result)
        return result

Now, thanks to Redis distributed locks, the function is always allowed to finish before it starts another one. All the hairy locking (in particular, the waiting) is implemented deep down in Redis which is rock solid.

Bonus Usage

Another use that has also saved my bacon is functions that aren't necessarily called with the same input argument but each call is so resource intensive that you want to make sure it only does one of these at a time. Suppose you have a Django view function that does some resource intensive work and you want to stagger the calls so that it only runs it one at a time. Like this for example:

def api_stats_calculations(request, part):
    if part == 'users-per-month':
        data = _calculate_users_per_month()  # expensive
    elif part == 'pageviews-per-week':
        data = _calculate_pageviews_per_week()  # intensive
    elif part == 'downloads-per-day':
        data = _calculate_download_per_day()  # slow
    elif you == 'get' and the == 'idea':
        ...

    return http.JsonResponse({'data': data})

If you just put @lock_decorator() on this Django view function, and you have some (almost) concurrent calls to this function, for example from a uWSGI server running with threads and multiple processes, then it will not synchronize the calls.

The solution to this is to write your own function for generating the lock key, like this for example:

@lock_decorator(
    key_maker=lamnbda request, part: 'api_stats_calculations'
)
def api_stats_calculations(request, part):
    if part == 'users-per-month':
        data = _calculate_users_per_month()  # expensive
    elif part == 'pageviews-per-week':
        data = _calculate_pageviews_per_week()  # intensive
    elif part == 'downloads-per-day':
        data = _calculate_download_per_day()  # slow
    elif you == 'get' and the == 'idea':
        ...

    return http.JsonResponse({'data': data})

Now it works.

How Time-Expensive Is It?

Perhaps you worry that 99% of your calls to the function don't have the problem of calling the function concurrently. How much is this overhead of this lock costing you? I wondered that too and set up a simple stress test where I wrote a really simple Django view function. It looked something like this:

@lock_decorator(key_maker=lambda request: 'samekey')
def sample_view_function(request):
    return http.HttpResponse('Ok\n')

I started a Django server with uWSGI with multiple processors and threads enabled. Then I bombarded this function with a simple concurrent stress test and observed the requests per minute. The cost was extremely tiny and almost negligable (compared to not using the lock decorator). Granted, in this test I used Redis on redis://localhost:6379/0 but generally the conclusion was that the call is extremely fast and not something to worry too much about. But your mileage may vary so do your own experiments for your context.

What's Needed

You need to use django-redis as your Django cache backend. I've blogged before about using django-redis, for example Fastest cache backend possible for Django and Fastest Redis configuration for Django.

This is an advanced topic for people who do serious stuff in Redis. I need to do serious stuff in Redis so I'm trying to learn about the best way to store lots of keys with hash maps.

It seems that this article by Salvatore Sanfilippo (creator of Redis) himself seems to be a much cited article for this topic. If you haven't read it, the gist is that Redis can employ some clever optimizations for storing hash maps in a very memory efficient way instead of storing each key-value separately.

"Hashes, Lists, Sets composed of just integers, and Sorted Sets, when smaller than a given number of elements, and up to a maximum element size, are encoded in a very memory efficient way that uses up to 10 times less memory (with 5 time less memory used being the average saving)"

This efficient storage optimization is called a ziplist.

How does that work?

Suppose you have 1,000 keys (with their values) if you store them like this:

SET keyprefix0001 valueX
SET keyprefix0002 valueY
SET keyprefix0003 valueN
...
INFO

then, the total memory footprint of the Redis DB is linear. The total memory footprint is roughly that of the length of the keys and values.
Alternatively, if you use hash maps (essentially storing dictionaries instead of individual values) you can benefit from some of those "tricks". To do that, you need to "bucketize" or "shard" the keys. Basically, some way to lump together many keys under one "parent key". (Note this is only applicable if you have huge number of keys)

In this article for one of the founders of Instagram, Mike Krieger, proposes a technique of dividing his keys (since they're all numbers) by 1000 and use the "remainder" as the key inside the hash map. So an original key like "1155315"'s prefix becomes "1155" (1155315 / 1000 = 1155) so he stores:

HSET "1155" "1155315" THEVALUE

Since the numbers are all integers the first 4 characters means there are 10,000 different combinations. I.e. a total 10,000 hash maps.

So when you need to look up the value of key "1155315" you apply the same logic you did when you inserted it.

HGET "1155" "1155315"

So this is where hash-max-ziplist-entries comes in. It's a configuration setting. In my Redis 3.2 install (both in an official Redis Docker image and on my Homebrew install) this number is 512. So what does that mean?

I believe, it means the maximum number of keys per hash map to benefit from the internal space optimization storage.

Real experiment

To demonstrate my understanding to myself, I wrote a hacky script that stores 1,000,000 keys with different "bucketization algorithms" (yeah, I made that term up). Depending how you run the script, it splits the keys (which are all integers) into different number of buckets. For example, if you integer divide each key by 500 you get 2,000 different hash map buckets, aka. total number of keys in the database.

So I run that script with a bunch of different sizes, flush the database between each run and at the end of every run I log how big the whole database is in number of megabytes. Draw a graph of this and this is what you get:

What's happening here is that since my hash-max-ziplist-entries is 512, around there, Redis does not want to store each hash map in space efficient way and stores it has plain key values.

In other words, when each hash map has more than 502 keys, memory usage shoots up.

Sanity check

It's no surprise that the more keys (and values) you store, the bigger the total size of the database. Here's a graph I used when inserting 10,000 integer keys, 20,000 integer keys etc.

The three lines represent 3 different batches of experiments. SET means it inserts N keys with the simple SET operator. HSET (1,000) inserts N keys with each key bucket is 1,000 keys. And lastly, HSET (500) means each hash map bucket has at most 500 keys.

I guess you could say they're all linear but note that they all store the exact same total amount of information.

What's the catch?

I honestly don't know but one thing is clear; there's no such thing as a free lunch.

In the memory-optimization article by Salvatore he alludes that "... a clear tradeoff between many things: CPU, memory, max element size.". So perhaps this neat trick that saves so much memory can comes at a cost of more CPU resource work. At the moment I don't know of a good way to measure this. So far my focus has been on the optimal way of storing things.

Also, I can only speculate but common sense smells like if that number (hash-max-ziplist-entries) is very large, Redis might allocate more memory for the tree it might store than it actually stores. Meaning if I set it to, say, 10,000 but most of my hash maps only have around 1,000 keys. Does that mean that the total memory usage goes up with empty memory allocations?

What to do

Why it's set to 512 by default in Redis 3.2 I don't know, but I know you can change it!

redis-store:6379> config get hash-*
1) "hash-max-ziplist-entries"
2) "512"
3) "hash-max-ziplist-value"
4) "64"
redis-store:6379> config set hash-max-ziplist-entries 1024
OK

When I make that change and run the original script again but using 1,000 keys per bucket, the whole thing only uses 12MB. So it works.

I haven't actually tested this in AWS ElastiCache but I did try creating a new ElastiCache Parameter Group and there's an input for it.

According to this page it seems it's possible to change it with "Azure Redis Cache" too.

Lastly, I wish there was a way to run Redis where it spits out in a log file somewhere whenever your trip certain thresholds.