40

Cloud Storage and Data Warehouse Destinations

How it works

Rather than streaming data to the destination in real time, like with most destinations, our data warehouse and storage destinations send data to your storage buckets in bulk at regular, 10-minute intervals. When we load data, we insert and update events, people, and groups, in JSON, CSV, or parquet files that we upload to your storage bucket. You can then ingest those files into the data warehouse or database of your choice.

These destinations only create new files in your storage bucket; they’ll never overwrite or append an existing file, so you can delete or remove files from your storage bucket after you ingest them into their ultimate destination—your data warehouse or database.

Exported files

Our data warehouse and cloud storage destinations generate parquet, JSON, or CSV files that we load in a storage bucket you specify. The data we send (the files we generate in your storage bucket) are based on the Actions you enable. But, by default, that means we generate files for all our standard source calls: identify, group, page, screen, track, and alias.

Files are named in the format workspaceName-dataType-syncNumber. For example, if your workspace is called production, files for track calls will be called production_tracks_1 where _1 is the first file.

Sync frequency

Unlike other destinations where we send data in real time, these kinds of destinations attempt to send data to your storage bucket every 10 minutes—though actual sync intervals and processing times may vary. When syncing large data sets, or when you have a high volume of concurrent sync operations, it can take a little longer to process and export data.

Each sync file contains data from the previous sync interval. For example, if the last sync occurred at 12:00 PM, the next sync will only send data from 12:00 PM to 12:09:59 PM.

Handling objects and arrays in CSV and Parquet files

Our source libraries pass nested objects and arrays into tracking calls as properties, traits, and tracking calls, but CSVs and Parquet files don’t have a concept of objects or arrays. So we stringify or flatten properties and traits in CSVs and Parquet files to preserve source data without significantly manipulating it.

{
  "received_at": "2019-08-24T14:15:22Z",
  "id": "a7280cfea0f6d",
  "user_id": "97980cfea0067",
  "anonymous_id": "d19b0cfeb606a",
  "sent_at": "2019-08-24T14:15:22Z",
  "traits": {
    "name": "Cool Person",
    "email": "cool.person@example.com",
    "likes_baseball": true
  },
  "context": {
    ...
  }
}

received_at,id,user_id,anonymous_id,sent_at,traits,context
2019-08-24T14:15:22Z,a7280cfea0f6d,97980cfea0067,d19b0cfeb606a,2019-08-24T14:15:22Z,"{\"name\": \"Cool Person\", \"email\": \"cool.person@example.com\", \"likes_baseball\": true}", "{...}"

Schemas

When we load data into your storage buckets, we create and update files to match the shape of your source data. Note that we flatten or stringify nested objects and arrays according to the rules above.

Identifies schema

Identifies files contain identify calls made from your sources. The context and traits in the schema below are objects in JSON. In CSV and parquet files, these columns contain stringified objects.

Groups schema

Groups files contain group calls made from your sources. If your integration outputs CSV or parquet files, the context and traits columns contain stringified objects.

Page schema

Pages contains entries for the page calls your sources send to Customer.io. If your integration outputs CSV or parquet files, the context and properties columns contain stringified objects. If your integration outputs JSON files, the context and properties columns contain objects.

Screen schema

Screens files contain entries for the screen calls your sources send to Customer.io. If your integration outputs CSV or parquet files, the context and properties columns contain stringified objects. If your integration outputs JSON files, the context and properties columns contain objects.

Track Schema

Tracks contains entries for the track calls you send to Customer.io. It shows information about the events your users perform.

If your integration outputs CSV or parquet files, the context and properties columns contain stringified objects. If your integration outputs JSON files, the context and properties columns contain objects.

Alias

The Alias schema contains entries for the alias calls you send to Customer.io. It shows information about the users you merge, with each entry showing a user’s new user_id and their previous_id.

Timestamps

We associate four timestamps with every source call: timestamp, original_timestamp, sent_at and received_at. All four timestamps pass through to your warehouse, and it may help to understand the purpose of each.

In general, you should use timestamp when you query for historical events and received_at for all other queries based on time.

timestamp is the UTC-converted timestamp set by the Customer.io library. If you import historical events using a server-side library, this is the timestamp you’ll want to reference in your queries.

original_timestamp is the original timestamp set by the source library when the event/source call is created. This timestamp can be affected by device clock skew. You can override this value by manually passing a timestamp in your source calls, which we map to the original_timestamp. Generally, this timestamp should be ignored in favor of the timestamp column.

sent_at is a UTC timestamp set when source libraries send calls to Customer.io. This timestamp can also be affected by device clock skew.

received_at is a UTC timestamp set by the Customer.io source API when we receive a payload from a source library. All tables use received_at as the sort key.

id

Each row in your database has an id which is equivalent to the messageId that our source libraries pass in source calls. This is a unique identifier associated with the row.

Sort Key

All tables use received_at as the sort key. Amazon Redshift stores your data on disk in sorted order according to the sort key. The Redshift query optimizer uses sort order when it determines optimal query plans.