Network Error Logging reports browsers sent

Failed requests to the organization's sites, from Network Error Logging reports.

A site that sends the Report-To and NEL headers gets a report from a Chromium browser for every request that failed, including pages that never loaded and so never ran the tracker. The answer counts them over the window, per bucket, per error type, per HTTP status code and per host and path.

Beside each count stands affected_visitors, the distinct visitors behind it, and per error type the percentiles of the time until the failure. A visitor is a non-zero stitch, so only sites with store_stitch count them, and meta.stitchless_site_ids names the sites that do not.

Every count is weighted by 1 / sampling_fraction, the share of failures the site's NEL header asks browsers to report. meta.sampling describes that failure sampling and not the application's rate: is_sampled is true when a report of the window was sampled below 1.0, and estimated_metrics then names the weighted fields. A status code of 0 means the request got no HTTP response.

Filters: the site scope, stitch, the report's own report_id, error_type, nel_class, status_code_range, host, path and page_url, and the edge-derived dimensions (country, asn, device_type, browser, os, bot_class, is_bot, is_datacenter, bot_name, bot_type, datacenter, ua_anomaly). exclude takes error_type, nel_class, host, path, page_url and the eight edge labels. Any other filter is a 422 naming it, since an empty answer would read as "no errors". An entry of paths is a host and a path value that together match exactly it. Without an error_type filter every type counts, abandoned included: a visitor who left while a request loaded, which every site has. exclude: {"error_type": ["abandoned"]} leaves it out of every count and the path list. nel_class has no default either.