KMDB

The km-db gem should be useful to KissMetrics (KM) users. Its aim is to efficiently process data obtained with KM's "Data Export" feature.

Its main feature is to import dumps directly from S3 into a SQL database, optimized for typical queries (in particular, partitioned along the time dimension).

Once imported, you can run complex queries against your visit history, for instance run multivariate analysis.

Beware though, KM data can be huge, and processing it is taxing!

Installing

If you want to run "just" KM-DB, you might want to just use the app.

Otherwise, add this to your Gemfile if you're using Bundler:

gem 'km-db'

Configuration

KMDB is configured through environment variables. We recommend storing this settings in a .env file if running locally, and using foreman to start KMDB commands with the environment set.

Preparing your database

KMDB requires a MySQL database (to store events, properties, etc) and a Redis store running (to store batch jobs and cache data).

Set the following:

  • DATABASE_URL (required), e.g. mysql2://km_db_test@localhost/km_db_test
  • KMDB_REDIS_URL [localhost], e.g. redis://localhost/14

Then run:

$ kmdb-flush

to prepare your database.

Optimizing your database

If your dataset is large (over 1 million events), KMDB can partition your database, i.e. transparently split large tables into smaller buckets of continuous time periods.

Set the following:

  • KMDB_MIN_DATE (required), e.g. '2014-01-01'
  • KMDB_MAX_DATE (required), e.g. '2016-01-01'
  • KMDB_DAYS_PER_PARTITION (required), e.g. '7'

Then run:

$ kmdb-partition

Notes:

  • MySQL only supports up to 1024 partitions.
  • You shoud aim for less than 1 million events per partitions for performance.
  • You should run this before importing data, but it's possible to re-run it. The MIN_DATE will be ignored, and partitions will be added up to the new MAX_DATE (if larger).

Importing data

KMDB will fetch JSON files form the S3 bucket where you instructed KissMetrics to back up your data, parse them, and store information in the database.

It does so using resque for high parallelism of the import process; in our experience, it's perfectly possible to import 100GB of data in a few hours.

Set the following:

  • RESQUE_WORKERS (1), number of worker nodes.
  • KMDB_MIN_REVISION (optional, default 1), first KissMetrics revision file you want to import.
  • KMDB_REVISION_LOOKAHEAD (10), how many revision files to check after the last known one
  • KMDB_BATCH_SIZE (100), how many events to process per batch (advisory, may be higher as an entire second's worth of events will always be processed in one batch to preserve ordering).
  • AWS_BUCKET (required), the name of the S3 bucket where the data is stored.
  • AWS_ACCESS_KEY_ID (required).
  • AWS_SECRET_ACCESS_KEY (required).

Ignoring some users

You may want to ignore all events and properties for certain users, for instance the administrative users of your site (or employees).

Simply add their identities to the ignored_users table before import.

Whitelisting events

It's typical to have some noisy and/or shorter-lived events sent to KissMetrics, e.g. for testing purposes or for temporary monitoring.

Should you only want to import certain events, add their names to the whitelisted_events table before starting import.

If the table is left empty, all events will be imported.

Dealiasing users

When KissMetrics finds a way to tie two user identities as being a single actual user, it stores an "aliasing" event. KMDB de-aliases users automatically during import, and will store all events and properties against a single user identity (one that's numeric if any, otherwise the lexicographically lowest).

Using imported data

Using SQL directly

KMDB tries to stay close to the KissMetrics data, leaving you to interpret it. As such, the main tables are unsurprisingly events and properties.

Here's a summary of the data model:

events has one row for each imported event:

events
id
t
n
user_id

properties has one row for each property ever set on events or users

properties
id
t
key
value
user_id
event_id

events.n and properties.key reference the id column of the keys table; this is done for performance reasons (event and property names are only stored once):

keys
id
string

KMDB also keeps the original user identities around in users, although you'll probably never need them:

users
id
name

as well as all aliasing events:

aliases
id
name1
name2

Using ActiveRecord

The KMDB module exposes four ActiveRecord classes: Event, Property, User are the main domain objects.

Key is used to intern strings (event and property names) for performance.

Please consult the source of these models for details.