Databricks Unity Catalog Migration Pilot

Pilots a Hive Metastore → Unity Catalog migration end to end: it audits which tables can migrate (and why the rest cannot), turns that audit into a dependency-ordered execution plan, traces UC's two-level permission model when access breaks, and picks an environment-isolation pattern for the one-metastore-per-region constraint.

Overview

The September 30, 2026 forcing function. Databricks is deprecating the legacy Hive Metastore. On the cutover date the hive_metastore catalog goes read-only — no new writes, no schema changes — and stays available only for read-through while you migrate. Everything that writes to HMS (jobs, DLT pipelines, dashboards, external BI) must be pointed at a Unity Catalog table before then, or it breaks. This is not optional and it is not fast: the blockers below (legacy storage schemes, DBFS-root managed data, DENY-based ACLs) each require a physical data move or a grant rewrite, so the work is measured in weeks, not an afternoon.

> ⚠️ Verify the exact date and scope against your workspace's current > deprecation notice (Databricks account console → Previews / Announcements). > Deprecation dates move; this skill's job is to get you ready well ahead of > whatever the live date is — treat 2026-09-30 as the planning deadline, not an > excuse to wait for it.

This skill does four things, in order, and each is deterministic where it can be:

1. Detect — scripts/audit-hms-readiness.py classifies every HMS table by its

storage URI into READY (cloud-native, UC-governable), BLOCKED (a named un-migratable condition), or ORPHAN (a dangling HMS row pointing at deleted storage). The LLM never eyeballs a URI; the script owns the verdict.

1. Plan — the migration-planner subagent turns the readiness CSV into a

per-table plan (SYNC / DEEP CLONE / rewrite / skip) ordered by dependencies so a view never migrates before its base tables.

1. Trace — the uc-permission-tracer subagent walks UC's two-level access

model to answer "why can't user X read this?" in one pass — replacing ~90 minutes of doc-spelunking per ticket.

1. Isolate — /uc-env-pattern-picker recommends one of four dev/test/prod

isolation patterns under the single-metastore-per-region limit.

It is architecturally distinct from the v1 databricks-upgrade-migration and databricks-migration-deep-dive skills: those narrate migration in prose. This one runs a real readiness audit, emits a dependency-ordered plan, and traces live permissions. Deep, load-on-demand knowledge lives in references/; the arithmetic-and-classification lives in scripts/.

Prerequisites

The three role grants below are the hard dependencies and the most common reason the pilot stalls mid-flow. The skill checks the role chain first (Step 1) and reports exactly what is missing before doing any work.

• Account-admin — required to enable the UC system schemas

(system.information_schema, system.access) the audit reads, and to create the metastore/assign it to workspaces. Enable at the account level, not per workspace.

• Metastore-admin — required to CREATE CATALOG, run GRANT on UC objects,

and create external locations + storage credentials. Migration is a sequence of metastore-admin operations.

• Cloud IAM read access (AWS s3/IAM, Azure, or GCP) — required to diagnose a

storage-credential failure (the D2 case: a table is cloud-native but the UC storage credential's IAM role cannot assume the bucket). Bash(aws:*) is allowed for read-only IAM inspection; the skill never mutates cloud IAM.

• Databricks CLI authenticated (databricks auth login, or the

DATABRICKS_HOST + DATABRICKS_TOKEN env pair) and jq for parsing. Enumeration and grant SQL run through the CLI Statement Execution API.

• DATABRICKS_WAREHOUSE_ID set to a running SQL warehouse — every

statement-execution call requires it.

Authentication. All auth comes from the environment: the CLI's DATABRICKS_HOST + DATABRICKS_TOKEN (or databricks auth login), and read-only cloud credentials from the standard provider chain (aws sts get-caller-identity to confirm). No secrets are hardcoded.

Instructions

The pipeline is verify role chain → detect readiness → plan → migrate → trace access → isolate envs. Do them in order; Step 1 is a hard gate.

Step 1: Verify the Role Chain (hard gate — fail fast, not mid-migration)

Before touching data, confirm the running principal holds the roles the migration needs. A missing role surfaces here, not three tables into a SYNC.

bash
# Am I authenticated, and as whom?
databricks current-user me | jq -r '.userName'

# Metastore-admin check: only a metastore admin can read every grant. If this
# errors with PERMISSION_DENIED, the principal is not metastore-admin.
databricks api post /api/2.0/sql/statements --json "$(jq -n --arg wh "$DATABRICKS_WAREHOUSE_ID" \
  '{warehouse_id:$wh, wait_timeout:"30s",
    statement:"SELECT 1 FROM system.information_schema.catalog_privileges LIMIT 1"}')" \
  | jq -r '.status.state, (.status.error.message // "ok")'

Report the role chain status plainly: account-admin (needed to enable system schemas), metastore-admin (needed for grants + external locations), cloud IAM read (aws sts get-caller-identity). If any is missing, name who must grant it and STOP — do not start an audit you cannot finish.

Step 2: Enable & Confirm the System Schemas

The readiness audit reads system.information_schema; permission tracing reads system.access. Those schemas are individually gated — enabling one does not enable the others — behind an account-level flag AND a metastore-admin grant chain. If they are not enabled, run the bundled idempotent enabler (account-admin auth; a workspace PAT is rejected up front) to enable each schema and grant a group USE CATALOG system + USE SCHEMA + SELECT in one pass:

bash
python3 "${CLAUDE_SKILL_DIR}/scripts/enable-system-schemas.py" \
  --account-id "$DATABRICKS_ACCOUNT_ID" --metastore-id "$METASTORE_ID" \
  --grant-to data-governance --dry-run   # drop --dry-run to apply

Then confirm a UC metastore is attached to this workspace (databricks metastores current) — without one there is nowhere to migrate to. The full two-layer access model (account-admin enables, metastore-admin grants, neither inherits SELECT) is in ${CLAUDE_SKILL_DIR}/references/system-tables-access-model.md.

Step 3: Detect — Run the Readiness Audit

Enumerate the HMS schema(s) and classify every table. The script owns the verdict; feed it either live (--live SCHEMA) or the enumerated rows as JSON.

bash
OUT="${OUT:-$(pwd)/uc-migration-out}" && mkdir -p "$OUT"

# Live enumeration (needs DATABRICKS_WAREHOUSE_ID + an authenticated CLI):
python3 "${CLAUDE_SKILL_DIR}/scripts/audit-hms-readiness.py" \
  --live sales --summary --out "$OUT/readiness-sales.csv"

Every row lands in one of three buckets — READY (cloud-native path, SYNC-able), BLOCKED (a named condition), ORPHAN (dangling HMS entry, cleanup not migration). When a BLOCKED class appears, load ${CLAUDE_SKILL_DIR}/references/uc-migration-blockers.md for that class's canonical error, the physical relocation procedure, and the per-cloud variant. Do not skip the CLONE-drops-history gotcha in that file — a shallow CREATE TABLE ... CLONE breaks time travel post-migration; use DEEP CLONE.

Step 4: Plan — Dependency-Ordered Migration Plan

Hand the readiness CSV to the migration-planner subagent (see agents/migration-planner.md) with the org's catalog naming convention. It emits a numbered execution order with per-step rationale and the right verb per table (SYNC for ready Delta externals, DEEP CLONE/rewrite for blocked, CREATE VIEW after base tables, skip for orphans). Dependency ordering is the point: a view never migrates before the tables it reads.

Invoke it via /audit-hms-readiness (which runs Step 3 then routes here) or directly by handing the subagent the CSV path.

Step 5: Migrate — Execute the Plan

Walk the plan top to bottom:

• READY (Delta external, cloud path): SYNC TABLE <uc_cat>.<schema>.<table>

FROM hive_metastore.<schema>.<table> — metadata-only, no data copy. SYNC SCHEMA does a whole schema when every table is ready.

• BLOCKED (legacy scheme / DBFS root): relocate the data first (DEEP CLONE

for Delta, CREATE TABLE AS SELECT for non-Delta) to a UC-governed cloud path, register an external location + storage credential, then create the UC table at the new path. Full procedure per class in uc-migration-blockers.md.

• LEGACY_TABLE_ACL: the data may be ready, but the DENY-based grants cannot

auto-map to UC's allow-only model — re-author them (Step 6 traces them).

Re-run Step 3 after a batch to confirm the migrated tables drop out of BLOCKED.

Step 6: Trace — Diagnose Access After Migration

When a user hits PERMISSION_DENIED or AccessDenied on s3://… post-migration, route to the uc-permission-tracer subagent (/trace-uc-permission <user> <error>). It walks UC's two levels — account-admin/metastore-admin status → group membership (including nested-group caveats) → catalog/schema/table grants → the external-location/storage-credential grant for a cloud AccessDenied — and returns a single actionable line: "user X needs group Y membership AND `GRANT SELECT ON <obj>` run by metastore admin W."

When the fix is "add the user to group Y" but the grant still does not apply, the cause is usually the Entra→Databricks SCIM bridge silently dropping nested-group membership — see ${CLAUDE_SKILL_DIR}/references/scim-bridge-patterns.md for the connector's direct-members-only limitation and the three workarounds.

Step 7: Isolate — Pick an Environment Pattern

For a fresh UC layout, run /uc-env-pattern-picker. It asks compliance, cost, and BI-tool questions and recommends one of four patterns for the one-UC-metastore-per-region constraint, emitting a matching Databricks Asset Bundle target stub. The four patterns, their cost models, and tradeoffs live in ${CLAUDE_SKILL_DIR}/references/uc-environment-isolation-patterns.md.

Output

• A readiness CSV ($OUT/readiness-<schema>.csv, in the working dir) — one row

per HMS table: table_name, storage_uri, scheme, migration_blocker, suggested_action, each table bucketed READY / BLOCKED / ORPHAN with a READY/BLOCKED/ORPHAN tally.

• A dependency-ordered migration plan (markdown from migration-planner) — a

numbered execution order, the verb per table (SYNC/DEEP CLONE/rewrite/skip), per-step rationale, and views ordered after their base tables.

• A permission trace (from uc-permission-tracer) — the exact missing group

membership + grant + who must run it, for a specific user + error.

• An environment-isolation recommendation — the chosen pattern, its cost model,

and a ready-to-paste DAB target stub.

• A role-chain status up front — account-admin / metastore-admin / cloud-IAM,

with the exact grant and grantor for anything missing.

Error Handling

| Error | Cause | Solution | |-------|-------|----------| | PERMISSION_DENIED on system.information_schema | Principal is not metastore-admin, or system schemas not enabled | Step 1/2: report the missing role; an account admin enables system schemas, a metastore admin grants access. Stop until resolved. | | No metastore attached to the workspace | UC not set up for this workspace | databricks metastores current; if empty, an account admin must create/assign a metastore before any migration. | | SYNC fails with an unsupported-location error | Table is on a legacy scheme (wasbs:///adl:///dbfs:/) | It is BLOCKED, not READY — relocate per uc-migration-blockers.md (DEEP CLONE to a cloud path), then create the UC table. | | Time travel broken after migration | Shallow CREATE TABLE ... CLONE was used | Use DEEP CLONE — shallow clone copies metadata only and drops Delta history. Re-clone with DEEP CLONE. See the gotcha in uc-migration-blockers.md. | | AccessDenied on s3://… after a READY table migrates | UC storage credential's IAM role cannot assume the bucket | Trace with /trace-uc-permission; verify the external location's storage credential role and the bucket trust policy with read-only aws calls (D2). | | DESCRIBE DETAIL returns no location | Orphaned HMS entry (deleted storage) or a view | The audit marks it ORPHAN — clean up or skip; never report it as a migration blocker. | | DATABRICKS_WAREHOUSE_ID unset | No SQL warehouse for statement execution | Set it to a running warehouse before Step 1. |

Examples

Example 1: "How do I migrate off the Hive metastore?"

Runs the full pipeline. Step 1 confirms metastore-admin, Step 3 audits the schema (12 READY, 5 BLOCKED, 2 ORPHAN), the planner emits a numbered order (SYNC the 12 ready Delta tables first, DEEP CLONE the 3 dbfs:/user/hive managed tables to s3://, re-CREATE the 2 Parquet externals, skip the 2 orphans, then the 4 views), and the skill walks the plan — surfacing the Sept-30-2026 deadline and the CLONE-history gotcha inline.

Example 2: "Which of my tables can't migrate, and why?"

Step 3 only. The audit CSV lists each BLOCKED table with its scheme and the one-line reason; the skill loads uc-migration-blockers.md for each class and gives the physical relocation procedure per cloud.

Example 3: "User can't read a table after we migrated it."

Routes straight to /trace-uc-permission alice@corp.com "PERMISSION_DENIED: SELECT on main.sales.orders". The tracer returns: "alice@corp.com is in no group with a grant; add her to `data-analysts` AND run `GRANT SELECT ON TABLE main.sales.orders TO data-analysts` as metastore admin — she is not account-admin, so the grant will not inherit."

Example 4: "How do I keep dev/test/prod separate under Unity Catalog?"

/uc-env-pattern-picker asks the compliance/cost/BI questions and, for a cost-sensitive team that accepts shared lineage, recommends single-metastore-catalog-per-env (bronze_dev/bronze_prod) with a DAB var.env-parameterized target stub — noting the lineage-cleanliness tradeoff and the multi-account alternative for hard isolation.

Resources

• ${CLAUDE_SKILL_DIR}/references/uc-migration-blockers.md — the taxonomy of un-migratable HMS conditions: canonical error, physical relocation procedure, per-cloud variant, and the CLONE-drops-history gotcha.
• ${CLAUDE_SKILL_DIR}/references/uc-environment-isolation-patterns.md — the four dev/test/prod isolation patterns under one-metastore-per-region, with cost models and a DAB target stub.
• ${CLAUDE_SKILL_DIR}/references/system-tables-access-model.md — the two-layer access model (account-admin enables, metastore-admin grants), per-schema enablement, and the manual permission traversal.
• ${CLAUDE_SKILL_DIR}/references/scim-bridge-patterns.md — the Entra→Databricks SCIM nested-group limitation and the three workarounds.
• ${CLAUDE_SKILL_DIR}/scripts/audit-hms-readiness.py — deterministic HMS-table readiness classifier (READY/BLOCKED/ORPHAN → CSV).
• ${CLAUDE_SKILL_DIR}/scripts/enable-system-schemas.py — idempotent system-schema enabler + group grant chain (account-admin auth precheck).
• ${CLAUDE_SKILL_DIR}/agents/migration-planner.md — turns the readiness CSV into a dependency-ordered plan.
• ${CLAUDE_SKILL_DIR}/agents/uc-permission-tracer.md — traces UC's two-level access model for a user + error.
• Databricks: Upgrade to Unity Catalog · SYNC command · UCX (Databricks Labs)