[receiver/sqlserver] Fix query sample dropping sessions blocked on schema locks

[splunk-shanu]

Description

The query sample template used CROSS APPLY sys.dm_exec_sql_text(r.plan_handle). For a session waiting on LCK_M_SCH_S (Schema Stability) or LCK_M_SCH_M (Schema Modification), sys.dm_exec_sql_text returns zero rows — because the plan-cache lookup itself needs schema stability that the wait is holding up. CROSS APPLY then eliminates the outer row entirely, so blocked sessions were silently omitted from db.server.query_sample events. Users lost visibility into all schema-related blocking (index rebuilds, ALTER TABLE, statistics updates during load).

This PR:

• Changes CROSS APPLY to OUTER APPLY on sys.dm_exec_sql_text(r.plan_handle) so rows are preserved when the TVF returns nothing.
• Adds OUTER APPLY sys.dm_exec_input_buffer(r.session_id, r.request_id) as a fallback source for statement_text. dm_exec_input_buffer reads directly from the client connection's TDS buffer and still works when plan_handle is unresolvable.
• Wraps statement_text extraction in COALESCE(SUBSTRING(...), ib.event_info, '') so statement-level granularity (via statement_start_offset / statement_end_offset) is preserved for healthy sessions and only falls back to whole-batch text from input_buffer when the plan-cache lookup fails.
• Makes the WHERE clause NULL-safe for both o.TEXT and ib.event_info so encrypted stored procedures and other zero-row scenarios are no longer silently dropped.

The same class of bug also affected encrypted procs (WITH ENCRYPTION), plans aged out of cache during long blocks, and pre-parse sessions — all fixed by the same change.

sys.dm_exec_input_buffer requires the same VIEW SERVER STATE permission the receiver already needs.

Link to tracking issue

Fixes #49983

Testing

• Live-reproduced against SQL Server 2022 with two scenarios:
  • Session B waiting on LCK_M_SCH_S blocked by Session A holding LCK_M_SCH_M → confirmed row now emitted.
  • Session B waiting on LCK_M_SCH_M blocked by Session A holding LCK_M_SCH_S (via long-running SELECT with HOLDLOCK) → confirmed row now emitted.
  • Healthy multi-statement batch → confirmed statement_text still returns the specific waiting statement (e.g. WAITFOR DELAY '00:05:00'), not the whole batch — statement-level granularity preserved.
• Updated existing golden test testQuerySampleQuery.txt to match the new query.
• Added new regression test TestQuerySampleQueryDetectsSchemaLockBlocking that asserts each invariant needed to prevent this class of bug:
  • must not use CROSS APPLY sys.dm_exec_sql_text
  • must use OUTER APPLY sys.dm_exec_sql_text(r.plan_handle)
  • must reference sys.dm_exec_input_buffer
  • WHERE must be NULL-safe for both o.TEXT and ib.event_info
  • statement_text must use a COALESCE(...) chain with ib.event_info fallback
• Verified the regression test fails against the pre-fix template with a clear diagnostic message pointing at the offending construct.
• Full go test ./... in receiver/sqlserverreceiver passes.

Documentation

No user-facing documentation changes. documentation.md unchanged (no metric/attribute additions).

Authorship

• I, a human, wrote this pull request description myself.

[opentelemetry-pr-dashboard]

Pull request dashboard status

Merged · refreshed 2026-08-05 06:45 UTC

Status above doesn't look right?

• Anything look wrong? Report it with what you expected; it helps us improve the dashboard.

[sv-splunk]

Looks good

[singhvibhanshu]

/rerun

[ChrsMark]

LGTM, approved by code-owner. Will merge.