Automatically index your Heroku Postgres database with Dexter

Your database is clever enough to index itself

Posted by Valérian Galliat on May 28, 2017

Dexter is a new tool that just came out with a mind-blowing idea: automatically figuring out the indexes you need based on your database logs.

Your database knows which queries are running. It also has a pretty good idea of which index is best for a given query.

Introducing Dexter, the Automatic Indexer for Postgres — Andrew Kane

You should read the above introduction if you want to learn more about Dexter itself.

Basically, it uses hypothetical indexes through the HypoPG extension to test if adding an index on a column would have improved a given query performance, based on the logs of your production Postgres server.

This article won’t cover in details Dexter itself, since the introduction is pretty complete. However it mentions that because Dexter needs the HypoPG extension, we can’t use it with Heroku or Amazon RDS.

Today, we’ll see how to use Dexter with Heroku (or RDS with a few tweaks).

The idea

Dexter only needs access to the database with HypoPG for testing the query plan with and without hypothetical indexes, to see if they would help speed up some queries. Unless you want to let it automatically add indexes to your production database, there’s no real need for it to run on directly on production. Only the production logs are important.

Here, we’re going to feed the production logs from Heroku Postgres to Dexter, but let it analyze the queries and hypothetical indexes on a local Postgres, running with Docker.

Postgres Docker image with HypoPG

We’re going to build the following Dockerfile to use Postgres (9.6 in our case) with HypoPG:

FROM postgres:9.6

RUN apt-get update && apt-get -y install wget build-essential postgresql-server-dev-9.6

RUN wget https://github.com/dalibo/hypopg/archive/1.0.0.tar.gz && \
    tar xf 1.0.0.tar.gz && \
    cd hypopg-1.0.0 && \
    make && \
    make install
docker build -t postgres-hypo .

Preparing the data directory

I assume you already have a local database with the same schema as the production database, let’s say in /usr/local/var/postgresql. Make sure Postgres is stopped, and make a copy to work on it with Dexter:

cp -r /usr/local/var/postgresql ~/dexter-postgres

We’re gonna extract the default configuration of the dockerized Postgres, because we don’t want to run it with your local config (which might listen on an UNIX socket instead of a TCP port, and use a different access configuration):

mkdir ~/postgres-default
docker run --name=postgres -v ~/postgres-default:/var/lib/postgresql/data postgres-hypo
^C
cp ~/postgres-default/*.conf ~/dexter-postgres

This will run the server on an empty data directory in ~/postgres-default, and let it initialize the default configuration. Then we kill it and copy the container configuration to the real data directory.

Run the server

docker run --name=postgres -v ~/dexter-postgres:/var/lib/postgresql/data -p 5432 postgres-hypo

This will mount your ~/dexter-postgres in the container data directory, and bind the 5432 port from the container to your host.

You can connect to it with:

docker run -it --rm --link postgres postgres-hypo psql -h postgres -U <user> <db>

Then run the following to enable HypoPG:

CREATE EXTENSION hypopg;

Running Dexter

Install Dexter with:

gem install pgdexter

Then pipe your production logs to Dexter, and connect it to your local copy:

heroku logs -a your-app -p postgres -t \
    | sed 's/^.*: \[[A-Z0-9_]*] \[[0-9-]*]//' \
    | dexter --log-level debug postgres://<user>@localhost:5432/<db> \
    | tee dexter.log

The sed part is to strip the prefix of Heroku Postgres logs that Dexter does not recognize; before stripping it, the logs looks like:

2017-06-28T15:10:02+00:00 app[postgres.13761]: [DATABASE] [7-1]  sql_error_code = 00000 LOG:  duration: 4822.012 ms  statement: BEGIN READ ONLY;
2017-06-28T15:10:02+00:00 app[postgres.13761]: [DATABASE] [7-2]       SELECT column
2017-06-28T15:10:02+00:00 app[postgres.13761]: [DATABASE] [7-3]         FROM table
2017-06-28T15:10:02+00:00 app[postgres.13761]: [DATABASE] [7-4]        WHERE condition

And we need it to be like:

sql_error_code = 00000 LOG:  duration: 4822.012 ms  statement: BEGIN READ ONLY;
     SELECT column
       FROM table
      WHERE condition

I also use --log-level debug so Dexter outputs the queries that would be optimized with a given index suggestion, so we can make more sense of why that index would help. It also outputs the query cost before and after the index which is very neat.

Finally, I tee the Dexter output to dexter.log (so I can see it live but also access it in that file later).

After a couple minutes, Dexter analyzed a bunch of tables and tried different hypothetical indexes based on the slow queries it saw through the production logs.

Here’s an example output for our case:

2017-06-30T11:38:58-04:00 Index found: tickets (reference)
2017-06-30T11:38:58-04:00 Query 018d8cd6e02ac3fe61bf30cf0ca8f4c8cfb809b0d8 (Cost: 122555.87 -> 16.08)

SELECT stuff FROM tickets WHERE reference = 'some-reference';

Dexter tells us that we should add an index on the reference columns of the tickets table, and that would change the cost of the given query from 122555.87 to 16.08.

Conclusion

Who else better than your database can index your database?

Even though this is not entirely true, and you should definitely review those found indexes and ask if it’s the right thing for your app and the queries you want to optimize for, having an automated tool suggesting indexes based on the queries you actually run in productions is of great value.

It could even make sense that at some point, database engines add a layer that manages indexes automatically, without the need for external tooling, by analyzing the queries you run the most and identifying the tables where you would benefit more from write or read performance.


Photo by Benjamin Pley.