We should minimize the volume of exposure events.

At the very least, exposure events should not be sent more than once per experiment per page view even if there are multiple logExposure() calls on a page. Ideally, however, exposure events would not be sent more than once per experiment per epoch.

Design of memory mechanism

@cjming wrote:

Based on the concluding comments in T414739: [Spike] Design memory mechanism for exposureLogged flags, I'll try to succinctly revise the proposal and pseudocode to where we landed.
To summarize, we want to implement a 2-tier architecture (for JavaScript client-side - PHP server-side will will remain 1 tier because of caching issues) for a versioned exposure ledger that minimizes redundant logging while preserves data integrity during experiment changes/updates.

Tier 1 - in-memory floor

• scope: single page view (JS) or HTTP request (PHP)
• guarantees no more than one exposure logging event per request/view

Tier 2 - best effort, optimal ceiling

• scope: browser tab/session per origin (JS only)
• deduplicates exposures across internal navigations and refreshes using mw.storage.session
• records 1 exposure event per session per origin (i.e. exposure on jawiki is invisible to an exposure on kowiki)

Remembering if exposure was logged on the web

In T414740: [Spike] Determine best storage mechanism to use for storing exposureLogged flags on the web, @cjming wrote:

Implementation:

• use versioned keys - hash experiment parameters to bypass old storage and trigger a fresh exposure event when the experiment config state changes
• progressive exposure suppression (JS only) - capture session starts with a 5 minute TTL for the first 10 exposure attempts (to pre-empt rapid navigation or refreshes), then pivot to a 24 hour TTL to suppress for the rest of the lifetime of the tab
• for the panic button scenario, introduce a global epoch variable? Server provides a timestamp to recover from systemic/pipeline issues - forces all users to re-log exposures regardless of storage state
  • update experiment config in ExperimentManager::getExperimentConfig() to return a new key called global_reset_epoch

For JS and PHP implementation proposals refer to T414740#11635559

Acceptance criteria

• Experiment#sendExposure() only produces 1 experiment_exposure event per experiment per epoch
  • Test Kitchen PHP SDK
  • Test Kitchen JS SDK
• Automated tests
• Update documentation (see the note about minimization of volume being a work in progress in https://wikitech.wikimedia.org/wiki/Test_Kitchen/Conduct_an_experiment#Exposure_logging)