migrate-glue-devendpoint-to-interactive-sessions

Migrate Glue Dev Endpoint to Interactive Session

Move a developer off a legacy Glue development endpoint (Glue 0.9/1.0, no console since March 2023, billed continuously) onto a Glue interactive session (Glue 2.0+, sub-minute startup, idle timeout, console support). Follows the official checklist: https://docs.aws.amazon.com/glue/latest/dg/development-migration-checklist.html

The migration is additive and reversible until the final step: the session is created and validated alongside the still-running dev endpoint. The dev endpoint is deleted only after the user confirms the session works.

Reference Documentation

• references/migration-mapping.md -- DevEndpoint field → session magic/API mapping, the access-method decision table, IAM two-principal setup, and code-compatibility notes for the 0.9/1.0 → 3.0+ jump.

Workflow

1. Inventory the Dev Endpoint and Route

1.1. Confirm inputs: dev endpoint name, region, AWS CLI profile. MUST NOT proceed without a dev endpoint name.

1.2. Call glue:GetDevEndpoint(EndpointName=<name>). Store as source. Record: RoleArn, GlueVersion, WorkerType, NumberOfWorkers (or NumberOfNodes), SubnetId, SecurityGroupIds, Arguments, ExtraPythonLibsS3Path, ExtraJarsS3Path, PublicAddress/PrivateAddress, Status.

1.3. Route on access method — ask the user how they use the dev endpoint today (or infer from source). This determines the session interface, per the official checklist. See the decision table in references/migration-mapping.md:

• SageMaker / Jupyter / JupyterLab notebook → Glue Studio notebook (upload .ipynb)
• Zeppelin notebook → convert to Jupyter, then Glue Studio notebook
• IDE (PyCharm / VS Code) → IDE integration for interactive sessions
• REPL → local aws-glue-sessions package
• SSH → no direct equivalent; use the Glue Docker image for local dev

1.4. Route on code compatibility (this is the hard gate): interactive sessions run Glue 2.0+ only — they cannot run 0.9/1.0. If the dev endpoint's scripts depend on 0.9/1.0-specific behavior (HDFS, YARN configs, Spark 2.x semantics, Python 2 syntax), the code MUST be upgraded to Glue 3.0+ first.

• If the user has 0.9/1.0-specific code → migrate the code using the glue-09-10-migration skill (its breaking-change catalogue covers Spark 2.x→3.x, Python 2→3, Scala, Log4j, Parquet timestamps) BEFORE provisioning the session.
• If the code is already version-agnostic PySpark/GlueContext → proceed directly.

2. Set Up IAM (two-principal model)

Interactive sessions need two principals (the #1 setup failure). See references/migration-mapping.md for exact policies.

• Runtime role: pass to CreateSession. Reuse the dev endpoint's RoleArn — it already has the right Glue job permissions and trusts glue.amazonaws.com.
• Client principal: the user/role running the notebook or CLI. MUST have permission to call session APIs (e.g. AWSGlueConsoleFullAccess) AND iam:PassRole on the runtime role. Verify before creating the session.

3. Provision the Interactive Session

Build session config by mapping source fields (full table in the reference):

3.1. Default: --glue-version 4.0, same WorkerType/NumberOfWorkers as the dev endpoint, --idle-timeout 30, Python 3. Override on request ("use Glue 3.0", "use 10 workers", "idle timeout 60").

3.2. Create via the user's chosen interface:

• Notebook (Glue Studio / Jupyter): set magics in the first cell — %iam_role, %glue_version, %worker_type, %number_of_workers, %idle_timeout, %connections — then run a cell to start the session.
• Headless / CLI (use for automated validation): glue:CreateSession with --id, --role, --command Name=glueetl,PythonVersion=3, --glue-version, --worker-type, --number-of-workers, --idle-timeout. Poll glue:GetSession until Status=READY (typically under 1 minute).

4. Validate Equivalence

4.1. Run the developer's actual Spark/PySpark code in the new session and compare output to the dev endpoint.

• Notebook path: run the migrated .ipynb cells; confirm no errors and matching results.
• Headless path: submit the code with glue:RunStatement(SessionId, Code), poll glue:GetStatement until Statement.State=AVAILABLE, then read the result from Statement.Output.Data.TextPlain (the exact JSON paths — do not guess). A Statement.Output.Status=error means the code failed — capture Statement.Output.ErrorName/ErrorValue.

4.2. PASS criterion: the session produces functionally equivalent output. Spark version (2.4.x → 3.3.x+) and Python version (3.6 → 3.10) WILL differ; that is expected and not a failure. On PASS → Step 5.

4.3. On failure, classify the error before doing anything else:

• CODE_COMPATIBILITY — a Glue 0.9/1.0 breaking change surfacing on Spark 3.x / Python 3.10. Signals: SyntaxError, ImportError/ModuleNotFoundError, No FileSystem for scheme: hdfs, untyped Scala UDF, Log4j 1.x classes, Parquet timestamp/rebase errors, Scala 2.11 binary incompatibility, AWS SDK 1.11 method not found. Fix path: use the glue-09-10-migration skill's breaking-change catalogue to diagnose and patch the code. Validate the patched code by re-running it in THIS interactive session via glue:RunStatement — do NOT create or run a Glue job for validation (that is glue-09-10-migration's native mechanism; here the session is the validation surface). Then re-run 4.1.

• MIGRATION_CONFIG_GAP — the code is fine but the session is missing something the dev endpoint provided. Signals: cannot reach a VPC data store (no Glue connection), ModuleNotFoundError for a lib that was in ExtraPythonLibsS3Path, ClassNotFoundException for a jar that was in ExtraJarsS3Path, catalog tables not found (--enable-glue-datacatalog not carried over). Fix path: patch the session config (add %connections/--connections, %additional_python_modules, %extra_jars, or the catalog arg) per the mapping in Step 3, recreate the session, and re-run 4.1.

• TRANSIENT — provisioning/throttling/internal errors with no user-code traceback. Fix path: retry 4.1 once.

4.4. Attempt budget (mirrors glue-09-10-migration): MAX_FIX_ATTEMPTS = 5. Each fix-and-rerun consumes one attempt. After each failure, compute a fingerprint (error_name + ":" + failing_symbol); if a fingerprint repeats (the same failure recurred after its fix), go to the Exit path immediately — do not burn the rest of the budget on a fix that is not converging. If the error matches no class in 4.3, go to the Exit path. On reaching 5 attempts without a PASS, go to the Exit path.

4.5. Exit path (validation cannot be made to pass): MUST NOT proceed to Step 5 and MUST NOT delete the dev endpoint — it is the rollback. Keep the dev endpoint running. Stop the session (glue:StopSession) to avoid cost, but leave it for debugging. Report classification, the error, what was tried, and manual next steps:

MIGRATION RESULT: VALIDATION FAILED — dev endpoint RETAINED
Dev endpoint: <name>  (Glue <ver>)  -> KEPT (rollback intact)
Interactive session: <id>  -> STOPPED (kept for debugging)
Halt reason: <BUDGET_EXHAUSTED|CYCLE_DETECTED|UNCLASSIFIED>
Failure class: <CODE_COMPATIBILITY|MIGRATION_CONFIG_GAP|TRANSIENT|UNCLASSIFIED>
Error: <ErrorName>: <ErrorValue>
Fixes attempted: [list]
Manual next steps: [list]

5. Decommission the Dev Endpoint (confirm first)

5.1. Present the validation result, then explicitly ask the user to confirm deletion of the dev endpoint — it is the rollback path. Ask a direct yes/no question naming the endpoint, for example: "Validation passed. Delete dev endpoint '<name>' in <region> now? Its definition will be backed up to S3 first. (yes/no)" MUST wait for an affirmative reply. Treat anything other than a clear yes (no answer, "later", "keep it", silence) as NO: skip deletion, leave the dev endpoint running, and tell the user it was kept and how to delete it themselves later. MUST NOT delete on assumption or to "save cost" without this explicit yes.

5.2. Back up the dev endpoint definition before deleting (MUST NOT skip). A deleted dev endpoint cannot be recovered, so persist its full config first. Write the source definition captured in Step 1.2 (the raw glue:GetDevEndpoint response) as JSON to S3 — default s3://<a bucket the runtime role can write>/glue-devendpoint-backups/<name>-<region>.json; ask the user for the bucket if none is evident. This is the rebuild record: it captures RoleArn, GlueVersion, WorkerType/NumberOfWorkers, SubnetId, SecurityGroupIds, Arguments, ExtraPythonLibsS3Path, ExtraJarsS3Path, and PublicKeys, so the endpoint can be recreated with glue:CreateDevEndpoint if the migration must be rolled back. Confirm the s3:PutObject succeeded before proceeding.

5.3. On confirmation and a confirmed backup: glue:DeleteDevEndpoint(EndpointName=<name>). Remove any endpoint-specific scaffolding the migration added (registered SSH public keys, temporary Elastic IPs).

5.4. Output the result:

MIGRATION RESULT: SUCCESS
Dev endpoint: <name>  (Glue <ver>)  -> DELETED
Definition backup: s3://<bucket>/glue-devendpoint-backups/<name>-<region>.json
Interactive session: <id>  (Glue 4.0)
Access method: <notebook|IDE|REPL|docker>
Code upgrade: <none | via glue-09-10-migration>
Validation: PASSED (output matched)
Idle timeout: <n> min

Gotchas

• Interactive sessions cannot run Glue 0.9/1.0. If the code is version-specific, upgrading it to Glue 3.0+ is a prerequisite, not optional. Delegate to glue-09-10-migration.
• Two IAM principals. Missing iam:PassRole on the client principal is the most common CreateSession/AccessDenied failure. The runtime role alone is not enough.
• Always set --idle-timeout. Dev endpoints never timed out and billed 24/7. A session with no idle timeout reintroduces that cost. Default 30 min.
• VPC access changes shape. Raw SubnetId + SecurityGroupIds become a named Glue connection referenced by %connections. There is no raw-subnet session arg.
• No SSH successor. If the developer relied on SSH into the endpoint, there is no session equivalent — route them to the Glue Docker image for local development.
• The C-extension limitation is lifted. Dev endpoints rejected C-extension Python libs (e.g. pandas); %additional_python_modules supports PyPI and S3 wheels.
• Decommission is destructive and last. Never delete the dev endpoint before the session is validated and the user confirms. It is the only rollback.
• Back up the dev endpoint definition to S3 before deleting. A deleted dev endpoint is unrecoverable; the saved JSON config is the only way to recreate it via glue:CreateDevEndpoint. MUST NOT skip, even when the user says "just delete it."
• DynamicFrame.fromDF emits a benign UserWarning on Spark 3.x ("DataFrame constructor is internal"). It is not an error; output is unaffected.

Troubleshooting

References

• references/migration-mapping.md — full config mapping, access-method decision table, IAM two-principal setup, code-compatibility notes.