[Done] Usage statistics

[jaychia]

Usage Statistics

To help developers understand how to improve Daft, we wish to collect statistics on how Daft is used, common bugs and performance bottlenecks.

This document describes a proposal for how collection of these statistics can be carried out in a way that is reviewable, easy to opt-out of, and unidentifiable for Daft users.

High Level

1. Reviewable: easy to review and delete any data that was collected by providing a Daft Session ID to Daft developers
2. Opt-out: Easy to opt-out using an DAFT_ANALYTICS_ENABLED=0 environment variable
3. Identification: Statistics are collected based on session keys which are automatically generated on every Daft import. These cannot be used to identify a person.

We will not:

1. Collect any of our users' proprietary code/data
2. Sell or buy any data about our users

What are we collecting?

We collect the following:

• Execution metadata: to understand common usage patterns and bottlenecks in Daft performance
• Daft API usage: to understand common usage patterns
• Daft internal error messages: to improve the stability of Daft

Execution metadata

• Daft Installation ID
• Daft version
• Daft git commit
• Operating system
• Python version
• Daft runner
  • Ray metadata (if on the Ray runner): version, cluster size, available CPUs, available memory, disk space
• Query plan execution metadata
  • Type of query node
  • Execution wall time
  • Memory usage

We do not capture any identifiable information in the query such as filepaths, column names or partition data.

Daft API usage

• Daft Installation ID
• DataFrame method calls
  • Count
  • Walltime
  • Method parameters are not logged

Daft internal error messages

• Daft Session ID
• System information: see execution metadata for more details
• Relevant stack trace: We do not collect stack traces in user-code. All stack traces will be filtered for stack frames arising from the getdaft module only.

[jaychia]

Implemented in v0.0.21