Category Archives: NoSQL

How we tweaked Postgres upsert performance to be 2-3* faster than MongoDB

As we all know, relational databases are fine if you’re dealing with small amounts of data but for web-scale high performance high inserts speed and masses of queries per second, NoSQL is what you need. At least, that’s the conventional wisdom/hype surrounding NoSQL databases such as MongoDB. However as we’ve recently discovered this is very wrong indeed.

As the conventional wisdom said that, we went with MongoDB for our first attempt at a system that needed to support a very high number of upserts – we were aiming for 50-100k upserted rows per second of unstructured data to be exact. Initial attempts with MongoDB on some very nice hardware (2Tb SSD SAN) produced about 10k upserts per second – good but not great. Then, this number increased when we started using batched upsert operations that was added in MongoDB 2.6, and increased even further with the advert of MongoDB 3.0 and the WiredTiger storage backend to 30-50k upserts/second. Very nice!

However, there are a number of issues with MongoDB in particular and limitations on NoSQL databases in general that relational (SQL) databases make easy – want to do aggregation, joins, distinct row selection from an index, complex data manipulations before transmitting data over the nework? Easy with SQL, virtually impossible with NoSQL. So, we started to think that if we could find a relational database with even vaguely good insert performance it would make life and development tasks so much easier that perhaps it would be worth the effort to change. As postgres generally seems to be able to handle anything you can throw at it, including support for a binary JSON data type we started looking there.

In order to avoid changing our database code very much so that we could compare as closely with MongoDB as possible, I created a really simple schema for the tables – the column(s) required for a primary key, and a jsonb data type column to store the record. One great feature of the recently released Postgres 9.5 release is the single statement upsert ability that reduces latency and code complexity, so all testing was done on this. As Postgres is designed from the ground up to be reliable and not lose data (unlike MongoDB which has fsync disabled by default), it is very slow if you are just doing single inserts. When you start batching these in to transactions of say 100 or 1000 entries you can achieve a higher throughput. If you want to compare it with MongoDB though, you can choose to disable the fsync at the end of every transaction by using the command

SET LOCAL synchronous_commit TO OFF

This means that in the event of a power failure you may lose some transactions that returned successfully, however you will never risk getting database corruption (unlike with MongoDB). By setting this one command you get performance in line with MongoDB – 50k upserts per second. However this was pushing 500Mb/sec of disk bandwidth which was saturating even the SSD array.

But Postgres has a few more tricks up its’ sleeve. One of the downsides of Postgres being incredibly protective of data is that it has a lot of writes – first it writes all commands to the ‘Write-Ahead Log’ (WAL) and then it updates the underlying data and indexes, and the WAL is where much of this additional disk IO was coming from. Starting in Postgres 9.1 however there is a feature called ‘unlogged tables‘ which avoids writing to the WAL, at the cost of loosing the table data after an unclean shutdown. Enabling this is as simple as changing your ‘CREATE TABLE’ statement to ‘CREATE UNLOGGED TABLE’. When I switched to using unlogged tables, disk IO went to virtually nil for most of the run (just at the end when it flushed the tables to disk did it dump several hundred MB of data – the database and indexes) and we saw a throughput of 85k upserts per second – roughly twice that of Mongo. The downside of this is that at an unclean shutdown, unlogged tables are automatically truncated as they may be in an inconsistent state. My proposal would be to allow a method to mark an unlogged table as read-only in order to ‘freeze’ its state hence negating the need to truncate after a crash; however apparently this would be a bit tricky to implement.

Because we want to do lots of searches with our data we heavily rely on indexes. On Postgres these appear to entail quite a bit more overhead than mongo. When I added 8 additional indexes to the unlogged table performance suddenly dropped from 85k upserts/sec to 35k upserts/sec. Fortunately there’s an easy solution to this in our case– create a table with just a single index for the primary key. Then create a ‘Materialized view‘ which is a clone of the table, and add the indexes to that. When your updates have finished, simply run “REFRESH MATERIALIZED VIEW”. This takes perhaps a minute to run on our data of 2m rows (as it is a single query, it is presently bounded to a single CPU), but it means that we get full upsert performance and the indexes then get updated inside a single transaction which should hopefully lead to less fragmentation and means better overall performance and less overall CPU usage at the cost of the materialized table lagging behind updated data by about 1 minute – an acceptable tradeoff in our case. By default “REFRESH MATERIALIZED VIEW” exclusively locks the materialized view while it is being update but you can fix this at the cost of some additional CPU cycles by changing that to “REFRESH MATERIALIZED VIEW CONCURRENTLY”. Your commands then look like:

-- Run these once:
CREATE MATERIALIZED VIEW real_table AS select * from unlogged_pk_only_table;
CREATE INDEX ON real_table (…);

-- Run this when real_table updates complete
REFRESH MATERIALIZED VIEW CONCURRENTLY real_table;

So we’ve now got performance twice the speed of Mongo without tweaking any postgres performance parameters – can we do better? It turns out that some basic settings help improve performance pretty significantly:

shared_buffers = 3GB # Give postgres lots more memory to work with
effective_io_concurrency = 8 # We are on a nice RAID SAN – we can do multiple reads/writes at once

The following setting is key as well – we don’t want to commit at the end of a transaction because frequently we have massive of transactions on the same table. Instead, wait a short amount of time and then commit with any other transactions on the same table in a single batch:

commit_delay = 100000 # 100ms

With these settings, an unlogged table with a single primary key and a jsonb column can run at 100k upserts/second, roughly three times the insert performance of MongoDB.

As I mentioned above, one of the issues with upserting lots of data into a single table was that we were seeing massive IO spikes on that. It turns out that this can be significantly reduced by changing the checkpointing settings too. There’s a small sentence in the manual that is key here: “To ensure data page consistency, the first modification of a data page after each checkpoint results in logging the entire page content” – basically if you checkpoint too frequently (and the default is after only 80mb of commands) after each checkpoint, if you are still updating lots of entries in the table, basically you entire table is rewritten to the WAL. To work around this (at the cost of causing longer recovery time after an unclean shutdown) we want to reduce the frequency of checkpoints with regard to the amount of data written (it will still write a checkpoint by default every 5 minutes anyway):

max_wal_size = 10GB
min_wal_size = 1GB

With these settings, even a standard logged table now runs at 80-90k upserts/sec and only outputs around 50Mb/sec of data.

TL;DR, one of the main areas of hype about NoSQL has been to do with the performance of inserting data. However, not only does Postgres 9.5 provide all the functionality of a mature and flexible SQL database, with a few tweaks it also handles upserts at two to three times the speed of MongoDB with better data consistency and lower memory and CPU usage.

Mongo DB Pain

In my latest contract we’re gathering a lot of data very quickly (100k inserts/sec would not be unreasonable) and trying to save it into a database. Before I started the contract the decision had been made to use Mongo Database for this very reason, so when I joined I was glad to finally get a chance to play with some NoSQL software – up until now I’ve just used MySQL and PostgreSQL. There’s some software and programming languages where the more you use them the more they pleasantly surprise you. AngularJS is one of those pieces of software where you start with basic templating and form-element binding but quickly discover there’s a mass of great plugins available. PostgreSQL is another case in point – I had been using MySQL quite happily for years but since starting to use PostgreSQL for some geolocation projects (with the excellent PostGIS extension), I’ve started using it by default for any new projects or designs. At the start it seems like just another database server with some geospatial additions, but then you think “Perhaps I can get the database to validate my input data before inserting”. No problem for Postgres you can specify a regexp constraint for example CHECK (mac ~ '^[0-9a-f]{12}$'::text) to validate a mac address (but you probably want to use the built-in type macaddr). Good luck with doing this in MySQL – by default it doesn’t even care about inserting the wrong data type into a column just silently messes it up… Or perhaps you want to import data from another data source into Postgres – again Foreign Data Wrappers (FDW) to the rescue connecting in to pretty much any other database or data source that you could want.

With Mongo however the more I’ve learnt of it the less impressed I’ve become. Sure, you can do some nice searches without needing schemas and it’s pretty high performance however once you go past simple store/retrievs there are some pretty basic issues that have not been addressed. Here are two I’ve come across in the past week:

Firstly, we have a big table (500m records) with data in a composite primary key (id, timestamp). We have the index on that, and we want to get a list of all the ids in the table. No problem you think – it’s a simple btree; just walk it and you’ll pull out the ids. And yes, it should be like this however there’s no easy way to do this in Mongo. The first thing you see in the documentation is that there is a .distinct() function or a .group() function. Great to get the distinct ids just run db.collection.distinct('id'). However rather than doing something sensible and returning the distinct fields with a cursor it tries to return them all in a single BSON object which is capped at 16mb size. So, if you have less then perhaps 1 million distinct ids in your table it would work, once you go over some magic threshold then your entire process needs redesigning.

Then you notice that in the latest versions of MongoDB, the .aggregate function can return a cursor. So, something like:

db.collection.aggregate( [ { $group: { _id: "$id" } } ], { cursor: {} } )

should do it. Technically speaking yes, however an open bug report from 4 years ago explains that actually the $group function can’t use indexes. If you add an explain: true to the aggregate you see that it does a full collection scan. But, perhaps if you are clever and force the sort like:

db.collection.aggregate( [ { $sort: { id: 1 } }, { $group: { _id: "$id" } } ], { cursor: {} } )

it will use the index, and indeed the explain output shows that it does. However even then it has to search through the whole table before it returns the first item for the cursor! So, if your distinct result is possibly more than 16mb of data in a large table there doesn’t seem to be any way to get the data in a timely manner from the database.

Another issue that I ran into and reported upstream is to do with very poor index selection on the aggregate. See the ticket for more details, but if you run an aggregate query on a collection with the foo column as a unique key:

db.col.aggregate([ { "$match" : { "foo" : "abcd" } }, { "$sort" : { "_id" : 1 } } ])

for some reason Mongo doesn’t follow the obvious and use the first item in the aggregate pipeline as the index. Instead, it chooses to use the _id index meaning that the $match does a full table scan to find the one result that could have been instantly looked up if it had used the correct index. The .find().sort() pipeline doesn’t have this issue.

There was also a fun issue I discovered whereby when you get a hot row in the database (for example a lot of processes trying to use $inc to update a counter) performance gets absolutely killed by the spinlocks on the data structure. Rather than 50k upsert/sec on the cluster when we had this hot row it went down to about 3k upserts/sec.

This is not to say that Mongo is a bad product, certainly looking at posts from a few years ago it seems that it has come on leaps and bounds since then. However it’s not the be all and end all that many NoSQL champions would claim. And I’m not convinced that the other NoSQL type databases would be all that much better – certainly for basic store/fetch operations but as you get closer to needing the functionality of a traditional RDBMS they are just too immature. Other issues for us recently have been lack of JOIN statements, the lack of schema that means a simple typo on a query can cause it to hang doing a full table scan for a key that doesn’t exist, etc etc. These are not problems with Mongo per-se but rather an issue with the incredible flexibility that NoSQL products offer.

Would I use Mongo again in other projects? Yes probably, but only if we had a requirement for very high performance or schema-less flexibility that something like PostgreSQL simply couldn’t provide.