How to report Gecko Telemetry in engine-gecko via Glean

In Gecko, the Telemetry system collects various measures of Gecko performance, hardware, usage and customizations. When the Gecko engine is embedded in Android products through any of the engine-gecko-* components of Android Components (there is one component for each Gecko channel), and the product is also using the Glean SDK for data collection, then Gecko metrics can be reported in Glean pings. This article provides an overview of what is needed to report any existing or new Telemetry data collection in Gecko to Glean.

Important

Every new or changed data collection in Firefox needs a data collection review from a Data Steward.

Overview

Histograms are reported out of Gecko with a mechanism called streaming Telemetry. This mechanism intercepts Gecko calls to tagged histograms and batches and bubbles them up through the the GeckoView RuntimeTelemetry delegate. The engine-gecko-* components provide implementations of the delegate which dispatches Gecko metrics to the Glean SDK.

Reporting an existing histogram

Exfiltrating existing histograms is a relatively straightforward process made up of a few small steps.

Tag histograms in Histograms.json

Accumulations to non-tagged histograms are ignored if streaming Telemetry is enabled. To tag a histogram you must add the geckoview_streaming product to the products list in the Histograms.json file .

Add Glean metrics to metrics.yaml

The Glean SDK provides a number of higher level metric types to map Gecko histogram metrics to. However, Gecko histograms lack the metadata to infer the Glean SDK destination type manually. For this reason, engineers must pick the most appropriate Gecko SDK type themselves.

Read more about how to add Glean SDK metrics to the metrics.yaml file in the Glean SDK documentation.

Important

Every new or changed data collection in Firefox needs a data collection review from a Data Steward.

Example: reporting CHECKERBOARD_DURATION

The first step is to add the relevant tag (i.e. geckoview_streaming) to the histogram’s products key in the Histograms.json file.

{
  "CHECKERBOARD_DURATION": {
    "record_in_processes": ["main", "content", "gpu"],
    "products": ["firefox", "fennec", "geckoview", "geckoview_streaming"],
    "alert_emails": ["gfx-telemetry-alerts@mozilla.com", "kgupta@mozilla.com"],
    "bug_numbers": [1238040, 1539309],
    "releaseChannelCollection": "opt-out",
    "expires_in_version": "73",
    "kind": "exponential",
    "high": 100000,
    "n_buckets": 50,
    "description": "Duration of a checkerboard event in milliseconds"
  },
}

Note

Histograms with "releaseChannelCollection": "opt-in", or without a releaseChannelCollection specified in its definition are only collected on Gecko built for "nightly" and "beta" channels.

Since this is a timing distribution, with a milliseconds time unit, it can be added as follows to the metrics.yaml file:

gfx.content.checkerboard:
  duration:
    type: timing_distribution
    time_unit: millisecond
    gecko_datapoint: CHECKERBOARD_DURATION
    description: |
      Duration of a checkerboard event.
    bugs:
      - 1238040
      - 1539309
    data_reviews:
      - https://example.com/data-review-url-example
    notification_emails:
      - gfx-telemetry-alerts@mozilla.com
      - kgupta@mozilla.com
    expires: 2019-12-09 # Gecko 73

Please note that the gecko_datapoint property will need to point to the name of the histogram exactly as written in the Histograms.json file. It is also important to note that time_unit needs to match the unit of the values that are recorded.

Example: recording without losing process information

If a histogram is being recorded in multiple processes, care must be taken to guarantee that data always comes from the same process throughout the lifetime of a Gecko instance, otherwise all the data will be added to the same Glean SDK metric. If process exclusivity cannot be guaranteed, then a histogram (and the respective Glean SDK metric) must be created for each relevant process. Consider the IPC_MESSAGE_SIZE2 histogram:

{
  "IPC_MESSAGE_SIZE2": {
    "record_in_processes": ["main", "content", "gpu"],
    "products": ["firefox", "fennec", "geckoview"],
    "alert_emails": ["hchang@mozilla.com"],
    "bug_numbers": [1353159],
    "expires_in_version": "60",
    "kind": "exponential",
    "high": 8000000,
    "n_buckets": 50,
    "keyed": false,
    "description": "Measures the size of all IPC messages sent that are >= 4096 bytes."
  },
}

Data for this histogram could come, at the same time, from the "main", "content" and "gpu" processes, since it is measuring IPC itself. By adding the geckoview_streaming product, data coming from all the processes would flow in the same Glean SDK metric and would loose the information about the process it came from. This problem can be solved by creating three histograms, one for each originating process. Here is, for example, the histogram for the GPU process:

{
  "IPC_MESSAGE_SIZE2_GPU": {
    "record_in_processes": ["gpu"],
    "products": ["geckoview_streaming"],
    "alert_emails": ["hchang@mozilla.com"],
    "bug_numbers": [1353159],
    "expires_in_version": "60",
    "kind": "exponential",
    "high": 8000000,
    "n_buckets": 50,
    "description": "Measures the size of all IPC messages sent that are >= 4096 bytes."
  },
}

And the related Glean SDK metric

ipc.message:
  gpu_size:
    type: memory_distribution
    memory_unit: byte
    gecko_datapoint: IPC_MESSAGE_SIZE2_GPU
    description: |
      Measures the size of the IPC messages from/to the GPU process that are >= 4096 bytes.
    bugs:
      - 1353159
    data_reviews:
      - https://example.com/data-review-url-example
    notification_emails:
      - hchang@mozilla.com
    expires: 2019-12-09 # Gecko 73

The ipc.message.gpu_size metric in the Glean SDK will now contain all the data coming exclusively from the GPU process. Similar definitions can be used for the other processes.

How to access the data?

Once a new build of Gecko will be provided through Maven, the Android Components team will automatically pick it up. Because the Gecko train model has three channels, there are three engine-gecko-* components, one per Gecko channel: “engine-gecko-nigthly”, “engine-gecko-beta” and engine-gecko.

The availability of the metric in the specific product’s dataset depends on which channel the application is using. For example, if Fenix Release depends on the engine-gecko (release) channel, then the registry file additions need to be available on the Release channel for Gecko in order for them to be exposed in Fenix.

Unless Glean custom pings are used, all the metrics are reported through the Glean metrics ping.

Unsupported features

This is the list of the currently unsupported features: