MetricsBasedTesting
Launchpad Entry: karmic-qa-metrics-based-testing
Created: UDS Karmic
Contributors:
Summary
All our current testing is a binary pass/fail but we would also like to track the evolution of certain parameters such as boot speed and power usage that may not have a clear pass/fail threshold but for which historical data is desireable. We will extend out infrastructure to collect and display such data.
Rationale
Work to improve performance requires access to rich performance data rather than binary pass/fail.
User stories
Design
Gather non-binary data during various checkbox tests and feed them on to the developers who need this for their work.
Examples
- Tracking bootspeed (using bootchart)
- Power usage
- I/O throughput
- Amount of memory being used
- polled during install, application testing, immediately after boot, etc.
- Webcam framerate
Implementation
Initial pass should be able to be implented quickly and provide value to groups who need this data.
Phase I (Karmic)
- Gather metric data as attachments (files/blobs)
- Provide access to gathered metric data for developers
- Do not parse it in the certification system or results tracker -- just gather and regurgitate
Provide access to the data without doing any parsing
- Have a directory containing the data
- .../bootchart/$machinename/$timestamp.tgz or .../$machinename/metrics/bootchart/$timestamp.tgz
- linked from the certification 'librarian' so there is no necessity for copying the files around
- Have a file containing metadata about the machine from this view
- this data is already available from the website and can be retrieved programmatically
- information such as processor speed, ram, etc.
Phase II (Later)
- Store metric data in database as numeric value
- Do not yet parse or analyze heuristically
- Provide simple access for ad-hoc reporting
- Generate alerts when metrics change by a sufficiently large amount or above a threshold
Unresolved issues
- Should this data be public or private?
- Will proprietary data be exposed if the data is public? (E.g. unreleased hardware)
- We should be able to make a split the same way we do with the existing released data
- this is done with a flag in the certification database marking systems as private
- Should benchmarks be pigeonholed into the pass/fail/skip system?
- should numeric data be an addition to the existing result type or a new one?
- heno feels that this should be a new type -- our existing results ain't broke so don't fix 'em
- also certain data doesn't really necessarily have a pass or fail... e.g. boot speed
- should numeric data be an addition to the existing result type or a new one?