Custom SQL enrichment
Summary​
The SQL Query Enrichment lets you perform dimension widening on a Snowplow event via your own internal relational database.
Overview​
If you have data points that you’d like to use to enrich your event data collected with Snowplow that live in a data base, this enrichment will help you to query for the fields you want to add.
Currently supported database types:
- MySQL, plus variants which speak MySQL (e.g. MariaDB, Amazon Aurora)
- PostgreSQL, plus variants which speak PostgreSQL
We don’t recommend to use this enrichment with analytical databases which support minimal (50-100) concurrent queries (e.g. Redshift).
For help with configuring this enrichment and getting it live on your pipeline please contact us at support@snowplowanalytics.com.
Configuration​
Unsure if your enrichment configuration is correct or works as expected? You can easily test it using Snowplow Micro on your machine. Follow the Micro usage guide to set up Micro and configure it to use your enrichment.
Hypothetical example​
Below you can see an example configuration using imaginary PostgreSQL database with CRM data, used to widen Snowplow event with context containing information about users.
The configuration JSON for this enrichment contains four sub-objects:
inputs
 specifies the datapoint(s) from the Snowplow event to use as values to substitute placeholders inÂquery
 when performing SQL querydatabase
 defines how the enrichment can access your relational databasequery
 defines prepared SQL statement to query your databaseoutput
 lets you tune how you convert the returned row(s) into one or more self-describing JSONs ready to be attached to your Snowplow eventcache
 improves the enrichment’s performance by storing rows retrieved from your relational database
{
"schema": "iglu:com.snowplowanalytics.snowplow.enrichments/sql_query_enrichment_config/jsonschema/1-0-0",
"data": {
"name": "sql_query_enrichment_config",
"vendor": "com.snowplowanalytics.snowplow.enrichments",
"enabled": true,
"parameters": {
"inputs": [
{
"placeholder": 1,
"pojo": {
"field": "user_id"
}
},
{
"placeholder": 1,
"json": {
"field": "contexts",
"schemaCriterion": "iglu:com.snowplowanalytics.snowplow/client_session/jsonschema/1-*-*",
"jsonPath": "$.userId"
}
},
{
"placeholder": 2,
"pojo": {
"field": "app_id"
}
}
],
"database": {
"postgresql": {
"host": "rdms.intra.acme.com",
"port": 5439,
"sslMode": true,
"username": "snowplow_enrich_ro",
"password": "1asIkJed",
"database": "crm"
}
},
"query": {
"sql": "SELECT username, email_address, date_of_birth FROM tbl_users WHERE user = ? AND application = ? LIMIT 1"
},
"output": {
"expectedRows": "AT_MOST_ONE",
"json": {
"schema": "iglu:com.acme/user/jsonschema/1-0-0",
"describes": "ALL_ROWS",
"propertyNames": "CAMEL_CASE"
}
},
"cache": {
"size": 3000,
"ttl": 60
}
}
}
}
Configuration​
inputs
​
Specify an array of inputs
 to put instead of placeholders in prepared statement. Each input consists of a placeholder
 and a source: either pojo
 if the datapoint comes from the Snowplow enriched event POJO, or json
 if the datapoint comes from a self-describing JSON inside one of the three JSON fields. The placeholder
 is the index of the ?
 SQL placeholder which this input will be used to populate. It must be an integer greater than or equal to 1. For the pojo
 source, the field name must be specified. A field name which is not recognized as part of the POJO will be ignored by the enrichment. For the json
 source, you must specify the field name as either unstruct_event
, contexts
 or derived_contexts
. You must then provide two additional fields:
schemaCriterion
 lets you specify the self-describing JSON you are looking for in the given JSON field. You can specify only the SchemaVer MODEL (e.g.Â1-*-*
), MODEL plus REVISION (e.g.Â1-1-*
) or a full MODEL-REVISION-ADDITION version (e.g.Â1-1-1
)jsonPath
 lets you provide the JSON Path statement to navigate to the field inside the JSON that you want to use as the input
The lookup algorithm is short-circuiting: the first match for a given key will be used.
database
​
The database
 section lets you configure how the enrichment should access your relational database. At the moment postgresql
 and mysql
 are supported, with the same fields available for both. Please see the Database support section above for notes on compatibility. Populate all properties in the postgresql
 or mysql
 configuration object, as follows:
host
, the hostname or IP address of the database server or clusterport
, the port to connect to the database onsslMode
, whether the database requires connections to use SSLusername
, the username to login to the database usingpassword
, the password for this usernamedatabase
, the name of the specific database to run the query against
We strongly recommend that the username have minimal read-only permissions on just the entities required to execute the SQL query. If your database server has additional authentication in place, such as IP whitelisting, you will need to configure this security to permit access from all of your servers running the Snowplow Enrichment process.
query
​
Consists of a single sql
 key with SQL query in a form of prepared statement to run on your relational database to return row(s) to attach to this event. Form of a prepared statement means it can has placeholders (?
), which will be replaced with actual values, extracted from input
s corresponding to their indexes. Some notes on the behavior:
- If a placeholder index required in theÂ
sql
 prepared statement is not found in any of theÂinputs
, then the lookup will not proceed, but this will not be flagged as a failure - A finalÂ
;
 is optional - Values will be inserted into prepared statement according to their type: strings will be quoted, number won’t
- This enrichment makes no attempt to sanitize the SQL statement, nor to verify that the SQL statement does not have harmful side effects (such as SQL injection)
output
​
The enrichment adds the returned row(s) into the derived_contexts
 field within a Snowplow enriched event. Because all JSONs in the derived_contexts
 field must be self-describing JSONs, use the schema
 field to specify the Iglu schema URI that you want to attach to the event. The expectedRows
 enum defines expected SQL output and can take the following values:
EXACTLY_ONE
 – exactly one row is expected. 0 or 2+ rows will throw an error, causing the entire event to fail processingAT_MOST_ONE
 – either one or zero rows is expected. 2+ rows will throw an errorAT_LEAST_ZERO
 – between 0 and N rows are expected – in JSON terms we are dealing with an array of resultsAT_LEAST_ONE
 – between 1 and N rows are expected – i.e. an array of results. 0 rows will throw an error
It is on user’s behalf to make sure SQL query returns correct amount of rows. The describes
 property dictates whether the schema
 is the self-describing schema for all rows returned by the query, or whether the schema
 should be attached to each of the returned rows:
ALL_ROWS
 means that theÂschema
 should box all returned rows – i.e. one context will always be added toÂderived_contexts
, regardless of how many rows that schema containsEVERY_ROW
 means that theÂschema
 should be attached to each returned row – so e.g. if 3 rows are returned, 3 contexts with this same schema will be added toÂderived_contexts
The propertyNames
 property supports reformatting of the returned columns to fit the JSON Schema’s conventions better. Supported options are:
AS_IS
 – preserve the column names exactly as they areCAMEL_CASE
 – soÂdate_of_birth
 becomesÂdateOfBirth
PASCAL_CASE
 – soÂdate_of_birth
 becomesÂDateOfBirth
SNAKE_CASE
 – soÂdateOfBirth
 becomesÂdate_of_birth
LOWER_CASE
 – changes all characters to lowercaseUPPER_CASE
 – changes all characters to uppercase
If these options aren’t bespoke enough, remember that you can use column aliasing in your SQL statement to tweak individual column names.
cache
​
A Snowplow enrichment can run many millions of time per hour, effectively launching a DoS attack on a data source if we are not careful. The cache
 configuration attempts to minimize the number of lookups performed. The cache is an LRU (least-recently used) cache, where less frequently accessed values are evicted to make space for new values. For the cache key, we use a complex object with an underlying Indexed HashMap, consisting of placeholder numbers as keys and extracted values as HashMap values. You can configure the cache
 as follows:
size
 is the maximum number of entries to hold in the cache at any one timettl
 is the number of seconds that an entry can stay in the cache before it is forcibly evicted. This is useful to prevent stale values from being retrieved in the case that your DB can return different values for the same key over time
ignoreOnError
​
When set to true
, no failed event will be emitted if the SQL query fails and the enriched event will be emitted without the context added by this enrichment.
Examples​
Single row​
With this configuration:
...
"query": {
"sql": "SELECT username, date_of_birth FROM tbl_users WHERE user = ?"
},
"output": {
"expectedRows": "AT_MOST_ONE",
"json": {
"schema": "iglu:com.acme/user/jsonschema/1-0-0",
"describes": "ALL_ROWS",
"propertyNames": "CAMEL_CASE"
}
},
...
And this query result:
SELECT username, date_of_birth FROM tbl_users WHERE user = 123;
| username | date_of_birth |
|----------|---------------|
| karl | 1980-06-12 |
This would be added to the derived_contexts
 array:
{
"schema": "iglu:com.acme/user/jsonschema/1-0-0",
"data": {
"username": "karl",
"dateOfBirth": "1980-06-12T00:00:00"
}
}
With this query result:
SELECT username, date_of_birth FROM tbl_users WHERE user = 123;
| username | date_of_birth |
|----------|---------------|
No context would be added to the derived_contexts
 array, but the event would continue processing. With this query result:
SELECT username, date_of_birth FROM tbl_users WHERE user = 123;
| username | date_of_birth |
|----------|---------------|
| karl | 1980-06-12 |
| mary | 1975-03-22 |
An error would be triggered, on account of the AT_MOST_ONE
 setting, and the event would be rejected.
Multiple rows​
With this configuration:
...
"query": {
"sql": "SELECT * FROM product WHERE category = ?"
},
"output": {
"expectedRows": "AT_LEAST_ZERO",
"json": {
"schema": "iglu:com.acme/products/jsonschema/1-0-0",
"describes": "ALL_ROWS",
"propertyNames": "AS_IS"
}
},
...
And this query result:
SELECT * FROM product WHERE category = 'homeware';
| SKU | prod_name |
|-----|-----------|
| 123 | iPad |
| 456 | Ray-Bans |
This single context would be added to the derived_contexts
 array:
{
"schema": "iglu:com.acme/products/jsonschema/1-0-0",
"data": [
{
"SKU": "123",
"prod_name": "iPad"
},
{
"SKU": "456",
"prod_name": "Ray-Bans"
}
]
}
If we change the configuration to:
...
"output": {
"expectedRows": "AT_LEAST_ZERO",
"json": {
"schema": "iglu:com.acme/product/jsonschema/1-0-0",
"describes": "EVERY_ROW",
"propertyNames": "LOWER_CASE"
}
}
...
Then two contexts would be added to the derived_contexts
 array:
{
"schema": "iglu:com.acme/product/jsonschema/1-0-0",
"data": {
"sku": "123",
"prod_name": "iPad"
}
}
{
"schema": "iglu:com.acme/product/jsonschema/1-0-0",
"data": {
"sku": "456",
"prod_name": "Ray-Bans"
}
}
Algorithm​
This enrichment uses any relational database to fetch data in JSON format. Here are some clues on how this enrichment will handle some exceptional cases:
- if provided JSONPath is invalid – all events attempted to being enriched will be sent toÂ
enriched/bad
- if more than one context (derived or custom) matchesÂ
schemaCriterion
 – first one will be picked, no matter if following have higher SchemaVer - if input’s value found more than in one sources – last one will be picked, so try to put more precise input last (for example to get longitude/latitude pair use data from IP Lookup enrichment first and GPS-derived longitude/latitude second)
- if any of input key wasn’t found – SQL query won’t be performed and new context won’t be derived, but event will be processed as usual
- if DB query wasn’t successful for any reason or timed-out – event will be sent toÂ
enriched/bad
 bucket - if DB connection was lost – event will be sent toÂ
enriched/bad
 bucket, but on next event, enrichment will try to reestablish connection - all non-primitive values (objects, arrays) andÂ
null
s will be interpreted as not-found values
Data generated​
This enrichment adds a new context to the enriched event with this schema.
As during the SQL Query enrichment process the new context is added to derived_contexts
 of the enriched/good event, the data generated will end up in its own table determined by the custom schema
 key in output
 configuration sub-object.