Summary

• Fixes silent telemetry loss on SPOG (custom-URL) hosts when connecting to an all-purpose cluster via an http_path like sql/protocolv1/o/<workspace-id>/<cluster-id>.
• Session._extract_spog_headers now extracts the workspace ID from the /o/<wsid>/ path segment as a fallback when ?o=<wsid> is not present, and sets it as the x-databricks-org-id header so the workspace-scoped HTTP transports (telemetry, feature flags, SEA) can route correctly.
• Priority order preserved: explicit caller header ▶ ?o= in http_path ▶ /sql/protocolv1/o/<wsid>/ path segment.

Why

On a SPOG host the workspace identity has to be in either the URL or the x-databricks-org-id header so PoPP can route a request to the right workspace. For all-purpose cluster Thrift this is free — the workspace ID is in the /o/<wsid>/ segment of the http_path, so PoPP routes Thrift via routing_reason=workspace-id and the session opens fine without ?o=.

Connection-scoped HTTP clients used for telemetry, feature flags, and the SEA backend talk to different paths (/telemetry-ext, /api/...) that do not carry the workspace ID. The previous extraction only looked at ?o= in the query string, so on a cluster http_path without ?o= no x-databricks-org-id header was ever attached. PoPP fell back to default (account) routing on those endpoints and responded with a 303 redirect to /login — silently dropping every telemetry batch.

What changes

src/databricks/sql/session.py

• New compiled regex Session._CLUSTER_PATH_ORG_ID_RE matching (?:^|/)sql/protocolv1/o/(\d+)/[^/?]+.
• _extract_spog_headers now checks the caller-header guard first, then ?o=, then the cluster path segment. Empty http_path returns {} as before. The log line reports which source produced the workspace ID.

tests/unit/test_session.py

• Five new tests under the existing TestSpogHeaders class:
  • Cluster path without ?o= → header extracted from /o/<wsid>/
  • Cluster path with leading / → same
  • Cluster path with ?o= → query-param value wins
  • Cluster path + explicit caller header → caller wins
  • Warehouse path without ?o= → no header (regression guard: the new regex must not match warehouse paths)

Test plan

• pytest tests/unit/test_session.py::TestSpogHeaders -v — 12 tests pass (5 new).
• pytest tests/unit/test_session.py — full file passes (26/26).
• Behavior validated on the OSS JDBC equivalent fix against Prod SPOG (peco.azuredatabricks.net) all-purpose cluster: telemetry POST /telemetry-ext flipped from HTTP 303 → /login to HTTP 200 OK once the path-segment extraction populated x-databricks-org-id. Same shape of fix here.

Out of scope

• The corresponding OSS JDBC driver fix is opened as Extract SPOG org-id from cluster httpPath for non-Thrift requests databricks-jdbc#1472.
• The Node.js connector (databricks/databricks-sql-nodejs) already extracts the org ID from both query param and path segment (see extractWorkspaceId/hasMalformedOrgParam in lib/DBSQLClient.ts).
• The Go connector (databricks/databricks-sql-go) has a separate PR for the same fix.
• The C# ADBC driver (adbc-drivers/databricks) has the same parsing limitation in PropertyHelper.ParseOrgIdFromProperties and needs a corresponding fix in its own repo.

This pull request and its description were written with assistance from Claude Code.