[WIP] - Upstream 16530 - Optimize HostList API: conditional DISTINCT + composite index on JobHostSummary

[cigamit]

Upstream Summary

• Make .distinct() conditional on host_filter query parameter being set — without it, the RBAC IN subquery on a direct FK (inventory_id) cannot produce duplicates, so DISTINCT is pure overhead forcing PostgreSQL to sort/hash-dedup the entire result set
• Add composite index (host_id, id DESC) on main_jobhostsummary so the with_latest_summary_id() correlated subquery can use an index-only top-1 scan instead of scanning and sorting per row

The host_list_rbac query pattern is the #2 DB time consumer in Scale Lab testing at 2,871 seconds total, and is worsening (+15%). Unlike AAP-81082, the RBAC subquery itself is clean (simple IN on a direct FK) — the cost is dominated by the unconditional DISTINCT and the correlated subquery lacking a composite index.

Why .distinct() is safe to make conditional

SmartFilter.query_from_string() can filter through M2M relationships (e.g., groups__name=...), which produce duplicate rows via JOINs. .distinct() is only needed when host_filter is set. Without host_filter, the queryset flows through HostAccess.filtered_queryset() which filters on inventory_id IN (RBAC subquery) — a direct FK that cannot produce duplicates.

Why the composite index helps

with_latest_summary_id() adds a correlated subquery:

SELECT id FROM main_jobhostsummary WHERE host_id = <outer.id> ORDER BY id DESC LIMIT 1

This runs for every result row. The existing auto-index on host_id alone requires scanning all matching rows then sorting. The composite (host_id, id DESC) index enables a single backward index scan to fetch the top-1 result directly.

EXPLAIN analysis (Scale Lab, Jul 2)

Ran on the aap26-next read replica (37,824 hosts, 16.7M job host summaries, 320K role evaluations).

The correlated subquery is the smoking gun — it scans the entire 367MB PK index backward with a filter instead of using a targeted index:

Index Scan Backward using main_jobhostsummary_pkey (cost=0.43..624,118)
      Filter: (host_id = $0)

Cost 624,118 per host row. With ~23K summaries per host on average (max 92K), each probe is extremely expensive.

Query variant	Estimated cost	Delta
RBAC subquery alone	3,113	—
Minimal (no DISTINCT, no subquery, no JOINs)	3,131	baseline
Full query WITHOUT DISTINCT	7,376	+4,245
Full query WITH DISTINCT	7,384	+4,253

The correlated subquery adds +4,167 cost (56% of total). DISTINCT adds +8 in planner estimate but forces Unique + Sort on 138-byte rows.

Test plan

• All existing tests pass (210 host-related tests verified locally)
• EXPLAIN on Scale Lab confirms correlated subquery uses full PK scan (cost 624,118/row)
• EXPLAIN confirms no composite (host_id, id DESC) index exists
• Verify composite index is picked up by the correlated subquery after deployment
• Measure host_list_rbac mean query time improvement on Scale Lab

Classification

New or Enhanced Feature

Fixes: https://issues.redhat.com/browse/AAP-81517

[Copilot]

Pull request overview

Optimizes the HostList API query plan by reducing unnecessary DISTINCT usage and improving the correlated “latest JobHostSummary per host” lookup via a composite index on JobHostSummary.

Changes:

• Makes HostList.get_queryset() apply .distinct() conditionally instead of unconditionally.
• Adds a composite DB index on JobHostSummary (host_id, id DESC) to speed up the with_latest_summary_id() correlated subquery.
• Adds the corresponding Django migration for the new index.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
awx/api/views/init.py	Makes distinct() conditional in HostList.get_queryset() before annotating latest summary id.
awx/main/models/jobs.py	Adds a composite (host, -id) index on JobHostSummary to accelerate “latest summary per host” queries.
awx/main/migrations/0201_jobhostsummary_main_jobhostsumm_host_id_desc.py	Introduces the migration that adds the new composite index.