Reading results

yii3-ab-testing-clickhouse owns the analytics schema. Both delivery paths write into it; nothing in the family writes to it from PHP any more.

Tables

Table	Constant	Rows
ab_exposures_v2	AnalyticsSchemaV2::EXPOSURES_TABLE	one per exposure event
ab_conversions_v2	AnalyticsSchemaV2::CONVERSIONS_TABLE	one per conversion event, with goal and exposure_event_id

The columns are the canonical row from Events v2 minus the transport field v, plus ingested_at, which the table fills itself. Both tables are ReplacingMergeTree(ingested_at), partitioned by toYYYYMM(occurred_at) and ordered by (experiment, event_id).

AnalyticsSchemaV2::EXPOSURE_COLUMNS and CONVERSION_COLUMNS list the insert columns in order. They exist so producers can check themselves against the schema: an INSERT lists columns positionally, and a reordered producer would write a variant into the subject column without any error.

Creating them

use Rasuvaeff\Yii3AbTestingClickHouse\SchemaMigrations;

(new SchemaMigrations(client: $clickHouseClient))->apply();

apply() runs the shipped migrations through the toolkit's ClickHouseMigrationRunner and returns the names it applied — a second call returns an empty list. It is explicit on purpose: creating analytics tables is a deployment step, not a side effect of composer install. Pass other table names to apply() to rename them; SchemaMigrations::placeholders() gives the exact token map to an application that runs its own migration pipeline.

The v1 tables (ab_exposures, ab_conversions) are still created, so an installation upgraded from 1.x keeps reading its history. The two generations never share a table name and are not comparable row for row: v1 has no event identity and its timestamp is ingestion time.

Two non-negotiables

SELECT variant, uniqExact(subject_id) AS subjects
FROM ab_exposures_v2 FINAL
WHERE experiment = 'checkout-button'
  AND decision_reason = 'assigned'
GROUP BY variant

• FINAL (or GROUP BY event_id). ReplacingMergeTree collapses duplicates on merge, not on insert, and delivery is at-least-once. Without it a retried delivery is counted twice.
• decision_reason = 'assigned'. Forced QA traffic and both kinds of fallback are recorded, but they are not experiment participation.

Conversion queries follow the attribution rules: link by exposure_event_id where present, otherwise by the window, and count first conversions only unless you are measuring a value.

Retention

Opt-in TTL templates ship in the package's retention/ directory. They are never applied automatically — an upgrade must never start deleting analytics data. Review the SQL, then apply it through your deployment process.

Personal data

subject_id is personal data when it is a user id. On the request side, the analytics allow-list keeps unlisted context attributes out of the rows entirely, and DbAssignmentStore::forget() erases a subject's sticky pins. Erasing already-delivered rows from ClickHouse is your retention process's job.