[TOC]
This covers how to review metrics code in the Chromium codebase.
During code review ensure the following -
-
Verify the histogram is created under the right histogram 'namespace'. (i.e if the histogram name is A.B.C, then A is the histogram namespace). If this is a new one, check if there a similar one that already exists?
-
Histogram owners match the histogram owners guidelines.
-
If generating multiple histograms programmatically or defining common set of histogram, guide them to use patterned-histograms.
-
Verify that expires_after is reasonable. CL author should be able to justify it. See guidance at on histogram-expiry.
-
Summary section should describe what is measured, when is it recorded and when not. The summary should generally describe a single emission of sample and not the statistics in aggregate.
-
If the histogram is recorded only for some platforms, then it should be included in the summary (unless part of its name).
-
Histogram and enum names don't include special characters besides dot, underscores or slashes.
- Verify that the histogram follows the UMA histogram design recommendation. If issues are found please link to docs in your review comments to increase their visibility.
-
If modifying an existing histogram, request that the histogram be renamed if its meaning has changed significantly. Common practices are to add suffix such as 2 to the name. When doing so, the existing entry for the histogram should be removed and an obsoletion message should be added following the guidance to point to the new histogram.
-
Re-numbering enum bucket values is not allowed as these break backward compatibility with respect to the data stored.
-
When modifying enums it is better to add new values to the enum instead of re-purposing existing enum values.
-
Modifying enum labels / summary is safe and allowed without review. However, if reviewing these changes make sure the semantic meaning of the bucket remain unchanged. e.g re-labeling 'Has Error' to 'Has Warning' should not be allowed, while re-labeling 'Has Error' to 'Has Error (e.g this and that error)' is okay.
-
Verify that histogram buckets are not a privacy risk. Some of the types forbidden are if the buckets are encoding page contents, URL, domain name, or is including any other type of personally identifying or sensitive information. If during review you are unsure then do not hesitate to request that Chrome Privacy Team review the change.
-
Check that the histogram bucket space of all possible values for all clients will be limited to 50 generally. It should not exceed 100 unless the CL author has a justification for having extra buckets.
-
If the lower bucket counts prove insufficient (e.g after reviewing the data for 50 buckets proves to be unhelpful, more buckets can be requested in a separate review).
-
When logging hashes in a sparse histogram make sure the global space meet this criterion (not just a client). Another valid case is when logging API error codes which in theory could be potentially large, but on a particular day it is not the case.
-
-
Check that recorded samples will be within 0 and 2^31-1.
-
If you expect majority of values to be under a certain value, say X, then it's good practice to pick the next order of magnitude - e.g. 10X for the max. This ensures that even outliers will not overflow the distribution.
-
Verify that for an enum histogram, the enum described in enums.xml and the enum defined in the client code matches. Furthermore, the enum in the code should have a comment mentioning that the values must not be changes and also that additions to the enum should be synced to enums.xml.
- See the sample comment here. The comment should be identical or nearly so.
-
If the histogram is logged via a macro (rather than a call to the function), check that the names will be constant at runtime.
-
Metrics team members need not be versed with code where histogram is recorded and are not reviewing all the corner cases associated with collecting the histogram.
-
enums.xml changes don't need a review. However, it is still useful to verify that the changes match the guidelines mentioned above.
-
Verify that the user action logged is actually user triggered. If they are not then advise cl author to convert them to a histogram.
- However if ordering of actions is the interesting part of their analysis then maybe see if each order combination can become a histogram bucket instead of user action.
-
Don't allow logging of noisy user actions (like scroll events). Typical allowed frequency is to be less frequent than PageLoad or MobilePageLoaded event.
-
UKM metrics are to be reviewed by UKM data privacy owners.
-
The metrics must follow the data collection guideline.
The ChromeUserMetricsExtension proto includes a variety of other fields such as records for Omnibox, Profiler, Stability, etc. These are specialized reviews and should be routed to relevant owner. The guidelines here don't cover these cases and typically require a server-side review first to change the proto.