[ryanmrodriguez]

Problem

On the Jobs page, class search was a case-sensitive prefix match and the queue filter needed the exact queue name. Typing Docu found nothing, even with dozens of failed CoreModels::Screenshots::UserDocumentScreenshotJobs on screen.

Change

Both fields are now a case-insensitive substring match: charge finds Billing::ChargeSubscriptionJob, HOOK finds webhooks.

The obvious fix, LIKE '%frag%', can't use an index, so every list, all seven tab counts and bulk "all matching" would scan the jobs table. Instead, Flightdeck::KnownNames matches the fragment in Ruby against the distinct class and queue names (SELECT DISTINCT over the existing class_name / queue_name indexes), and JobsQuery filters with an indexed IN (...). A side benefit is that case folding is the same on every adapter (SQLite's LIKE folds ASCII only, MySQL's depends on collation).

Freshness and cost

• The name lists are memoized per process for one poll interval, independent of Rails.cache, so a filtered page costs a couple of DISTINCTs per interval even with no cache store.
• A search that matches nothing re-reads first (at most once a second), so a class that starts failing right after a deploy is findable immediately.
• Retry/discard run inside KnownNames.fresh { }, so "all matching" never skips a job with a brand-new class or queue name.
• Concurrent refreshes share one read (per-column mutex, connection checked out before the lock to avoid a pool deadlock), and the read bypasses the query cache.

The exact class_name param filter is unchanged.

Known limitation: on accent-insensitive MySQL collations, IN can match collation-equivalent names (e.g. email vs émail). The old exact queue filter behaved the same way.

Testing

• New test/known_names_test.rb: query count with no cache store, TTL expiry, no-match re-read and its rate limit, slow reads, contended threads sharing one read, query cache bypass, nested/raising fresh.
• Bulk retry/discard regression tests for brand-new class and queue names (pinned clock). Each new guard test was confirmed to fail without its fix.
• bundle exec rake test and FLIGHTDECK_TEST_API_ONLY=1 bundle exec rake test: 298 runs, 0 failures each (SQLite locally; PostgreSQL/MySQL via CI). Rubocop clean.
• Checked by hand in bin/demo.