cshuo commented on code in PR #20243: URL: https://github.com/apache/hudi/pull/20243#discussion_r4387330858
########## hudi-lakehouse/charts/hudi-spark-connect/README.md: ########## @@ -0,0 +1,318 @@ +<!-- +Licensed to the Apache Software Foundation (ASF) under one +or more contributor license agreements. See the NOTICE file +distributed with this work for additional information +regarding copyright ownership. The ASF licenses this file +to you under the Apache License, Version 2.0 (the +"License"); you may not use this file except in compliance +with the License. You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +--> + +# Hudi Spark Connect endpoint + +Deploys Spark Connect 4.1.3 with Hudi 1.2.0, Scala 2.13, and Java 17. +Tables use an external Hive Metastore and S3-compatible storage. +See the [Dockerfile](../../images/spark-connect/Dockerfile) for dependency versions. +The delivery scope is interactive SQL reads and writes, not general Spark job +submission or a required DataFrame-write integration. + +## Local-dev quickstart + +Requires Docker, minikube, Helm, and kubectl. Run from the repository root: + +```bash +# Start MinIO, Hive Metastore, and Spark Connect +./hudi-lakehouse/scripts/quickstart.sh --spark-connect + +# Also write example data and query it through Spark Connect (requires Maven and a JDK) +./hudi-lakehouse/scripts/quickstart.sh --run-example --spark-connect +``` + +The example uses the existing Spark writer to populate and register `default.trips`, +then prints city counts: `chennai = 33`, `san_francisco = 34`, `sao_paulo = 33`. +The query runs through PySpark inside the endpoint pod; no host client is needed. +Omit `--spark-connect` to query through Trino instead. + +To rebuild the image and reload the running endpoint: + +```bash +./hudi-lakehouse/scripts/quickstart.sh --spark-connect --rebuild-images +kubectl -n hudi-lakehouse rollout restart deployment/hudi-spark-connect +kubectl -n hudi-lakehouse rollout status deployment/hudi-spark-connect +``` + +## Install against existing infrastructure + +Build an image and supply your metastore, warehouse, and storage credentials: + +```bash +./hudi-lakehouse/scripts/build-connect-image.sh --registry my.registry/team --push +helm upgrade --install hudi-spark-connect hudi-lakehouse/charts/hudi-spark-connect \ + --namespace hudi-lakehouse --create-namespace \ + --set image.repository=my.registry/team/hudi-lakehouse-spark-connect \ + --set catalog.metastoreUri=thrift://my-hms:9083 \ + --set catalog.warehouse=s3a://my-bucket/connect \ + --set storage.s3.existingSecret=my-s3-credentials \ + --wait --timeout 10m +``` + +The Secret must be in the same namespace and contain `accessKey` and `secretKey`. +For temporary credentials, also set `storage.s3.sessionTokenKey`. +Without a Secret, Hadoop uses its default credential provider chain. +Local-dev credentials are for development only. + +Other settings in [values.yaml](values.yaml): + +| Setting | Purpose | +|---|---| +| `storage.s3.endpoint`, `region`, `pathStyleAccess` | MinIO or custom S3 endpoints | +| `spark.cores` | Local worker threads | +| `spark.master` | Empty means `local[cores]`; opt in to a remote master explicitly | +| `spark.driverMemory` | JVM heap; keep the pod memory limit higher | +| `resources` | Driver pod CPU, memory, and optional ephemeral-storage requests/limits | +| `scratch.sizeLimit` | Optional disk-backed `/tmp` volume limit, e.g. `20Gi`; empty by default | +| `spark.extraConf` | Additional Spark properties; no credentials or overrides of chart-owned settings | +| `hudi.defaults` | Server-side `hoodie.*` defaults loaded from `HUDI_CONF_DIR/hudi-defaults.conf` | +| `serviceAccount.name`, `automountToken` | Existing driver account and token mounting; required for Kubernetes executors | + +By default the endpoint runs one shared driver in local mode, without separate executors. +Keep one replica. Restarting it ends active sessions; tables remain in the +external metastore and object store. + +Size `/tmp` for peak concurrent shuffle/spill and temporary write data, with headroom. +For example, `--set scratch.sizeLimit=20Gi` caps the driver's scratch volume; it does +not reserve disk. Set `resources.requests.ephemeral-storage` for scheduling and +`resources.limits.ephemeral-storage` for the pod budget, allowing for logs and the +writable container layer as well. Node disk pressure can evict the pod before the +volume reaches its limit. Scratch data is lost with the pod; this is not table storage. +In cluster mode, size executor scratch separately; this setting applies only to the driver. + +The service has no authentication or TLS. Keep it on a trusted network: +any client that can connect can read and write using the server's storage credentials. + +## Hudi writer defaults and table services + +The generic chart does not pin a table version. The local-dev values set: + +```yaml +hudi: + defaults: + hoodie.write.table.version: "8" + hoodie.write.auto.upgrade: "false" +``` + +This keeps new tables compatible with the local-dev Trino reader. Hudi 1.2.0 +otherwise defaults to table version 9. These are defaults, not enforcement: +explicit session/table settings can override them. Changing them does not downgrade +existing tables. Verify the versions supported by every reader before changing the pin. +The smoke test deliberately does not repeat the pin in `CREATE TABLE`. + +Use `hudi.defaults` for deployment-wide defaults, `TBLPROPERTIES` for table-specific +settings, and session `SET` for a writer operation. For example: + +```sql +CREATE TABLE orders (id INT, amount INT, ts BIGINT) USING hudi +TBLPROPERTIES (type = 'cow', primaryKey = 'id', preCombineField = 'ts', + 'hoodie.clustering.inline' = 'false'); + +SET hoodie.clean.automatic = true; +INSERT INTO orders VALUES (1, 100, 1); +RESET hoodie.clean.automatic; +``` + +Table properties and session settings take precedence over global defaults in +the Spark SQL write configuration; consult the relevant Hudi operation's settings +when overriding them. Quote Helm property values when exact numeric text matters. +Do not store credentials in `spark.extraConf` or `hudi.defaults`: both are ConfigMaps. +Configuration changes roll the driver and end existing sessions. +Hudi's defaults-file parser keeps backslashes literally but normalizes whitespace; +the chart rejects leading/trailing or repeated whitespace rather than silently +changing such values. This differs from Spark's Java-properties escaping. + +The endpoint does not schedule separate table-service jobs. Services enabled on a +SQL writer run according to Hudi's configuration in the shared Spark application. +In Hudi 1.2.0 inline compaction and inline clustering default to **off**; automatic +cleaning is enabled by default. Enabling inline compaction/clustering adds work to +the write path and can affect interactive-query latency. For example, a MoR writer +can set `hoodie.compact.inline=true` and `hoodie.compact.inline.max.delta.commits=2`. +For externally managed compaction/clustering, the operator must arrange both +scheduling and execution; this chart creates neither a scheduler nor a worker job. + +The caller/operator owns writer configuration, concurrency mode and lock provider +across **all** writers, including external Spark/Flink jobs. Separate Connect sessions +do not provide multi-writer coordination. This endpoint adds no locking policy, +workload isolation, or custom recovery guarantees. Embedded timeline-server reuse +(`hoodie.embed.timeline.server.reuse.enabled`) is not enabled here and needs separate +evaluation before changing its lifecycle in a long-running driver. + +## Kubernetes executor configuration (opt-in) + +The chart exposes the configuration needed for a shared driver with separate +executors. Local mode remains the tested default; rendering this configuration is +**not** evidence of a validated cluster deployment. + +```yaml +spark: + master: k8s://https://kubernetes.default.svc:443 + extraConf: + spark.executor.instances: "2" + spark.executor.cores: "1" + spark.executor.memory: 1g +serviceAccount: + name: spark-connect-driver + automountToken: true +``` + +Provision this ServiceAccount and its executor-management RBAC in the release +namespace first. The chart does not grant permissions. In Kubernetes client mode, +it advertises the driver's Pod IP and Pod name, selects the release namespace, +and defaults the executor image to the same Hudi-enabled image. The image delegates +executor startup to Spark's base entrypoint. Rebuild it when updating the entrypoint; +an older image will not provide that behavior. Executors receive the configured S3 +Secret references, including a session token when present, without embedding values +in Spark properties. The Hudi defaults file is loaded by the SQL driver; the resolved +write configuration is passed by Hudi to tasks. + +Before using this mode, validate driver/executor networking, resource capacity, +image availability/pull credentials, and storage access from executor pods. Use +`spark.kubernetes.executor.container.image` for a different compatible executor image, +and Spark's executor settings for pull secrets, placement, workload identity, or +credentials other than the chart's S3 Secret. Driver pod settings do not automatically +apply to executors. Keep executors in the release namespace when using its Secrets. +Dynamic allocation/autoscaling is not enabled by this example and needs separate +configuration and validation. Keep one driver replica; this is not a tenant-isolated service. + +## Connect and validate with ADBC + +Forward the endpoint for local clients: + +```bash +kubectl -n hudi-lakehouse port-forward svc/hudi-spark-connect 15002:15002 +``` + +| Client | URI | +|---|---| +| PySpark | `sc://localhost:15002` | +| Foundry ADBC | `spark://localhost:15002?api=connect&auth_type=none&tls=false` | + +ADBC requires `api=connect`; its default is Thrift. In-cluster clients can use +`hudi-spark-connect.hudi-lakehouse.svc:15002` without port-forwarding. +The ADBC driver is installed on the client, not the server. + +For the CRUD and restart test, prepare a Python 3.11 environment: + +```bash +python3.11 -m venv .venv-connect +. .venv-connect/bin/activate +pip install -r hudi-lakehouse/local-dev/example/connect-requirements.txt dbc==0.3.1 +dbc install 'spark=0.2.1' +``` + +Stop any manual port-forward before running the test; it manages its own: + +```bash +PYTHON="$VIRTUAL_ENV/bin/python" \ + ./hudi-lakehouse/local-dev/scripts/test-spark-connect.sh --restart-endpoint +``` + +The test creates a temporary table, checks INSERT/UPDATE/DELETE/MERGE, restarts the +endpoint, verifies persistence, and cleans up. It interrupts active sessions. +On failure, follow the printed cleanup instructions. + +Additional checks against a local-dev deployment: + +```bash +PYTHON="$VIRTUAL_ENV/bin/python" \ + ./hudi-lakehouse/local-dev/scripts/test-spark-connect.sh \ + --restart-endpoint --sql-matrix --freshness --trino +``` + +- `--sql-matrix`: CoW time travel and incremental TVF, partitioned CoW with HMS + sync, MoR snapshot/read-optimized reads and a completed inline compaction. Checks + stored table version 8 without statement-level version overrides; set + `EXPECTED_TABLE_VERSION=9` when deliberately testing a v9 deployment. +- `--trino`: implies `--sql-matrix`; compares CoW and partitioned-table rows through + an **already running** `hudi-trino` deployment (override `TRINO_DEPLOYMENT` if needed). + Connect-only quickstart does not install Trino. An unavailable/incompatible Trino + fails the check rather than skipping it. This tests Connect-created tables, not + just external-writer tables queried by both engines. +- `--freshness`: keeps one ADBC connection open while a separate Spark JVM commits + data, adds a column, and then recreates its test table. The supplied fixture creates + short-lived writer pods using the same pinned image and shared catalog/storage; + it is **not** another Connect session and does not test Spark 3.5/Flink compatibility. + Allow an additional 1 GiB memory request (1.5 GiB limit), CPU, and permissions to + create/delete pods and ConfigMaps. Only uniquely named test resources are touched. + +Freshness checks do not automatically refresh or reconnect. If they fail on stale +metadata, rerun with `--freshness --refresh` to test explicit `REFRESH TABLE` after +each external change. Record which mode passes for your version combination; +do not claim automatic visibility based on a refresh-assisted run. The fixture +does not alter shared example tables, and cleanup never uses `DROP DATABASE CASCADE`. +Other external writers can use the Python CLI's `--external-writer <command argv>` +hook; it receives the test database and `commit-schema` / `recreate` action. + +These acceptance paths require live Spark/ADBC (and optionally Trino). Offline +tests validate their assertions, configuration and cleanup, not engine behavior. +The full live SQL/Trino matrix and same-session freshness checks remain pending Review Comment: test running details: https://docs.google.com/document/d/1x9hmID10lgH941up6S6xIdIUz-oBUIG3wMpRP6URTnM/edit?usp=sharing -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
