Dedicated Trino
Ship a Trino cluster as part of the Data Product. Vulcan provisions and owns one cluster per product.
Use this path when the Trino cluster should be part of the data product deployment itself. You don't manage the cluster separately from the product: a single vulcan trino resource provisions the coordinator and workers, generates catalogs from your depots or DataOS-provisioned secret catalogs, and runs the plan and run workflows.
This page covers the connection and deploy shape. For tuning, failure modes, and local development, see the Trino engine manual in References.
How it works
One vulcan trino resource creates 5 cooperating resources:
<name>-trino Trino coordinator (replicas: 1)
<name>-trino-workers Trino workers
<name>-plan vulcan migrate, then plan --auto-apply (waits for the cluster)
<name>-run vulcan run (scheduled; depends on plan)
<name>-api Vulcan API for consumersCheck their status with:
dataos-ctl resource -t vulcan -n <name> get runtime -rVulcan generates catalogs automatically: one per depot in deployment.depots[], plus one per secret catalog you define.
Prerequisites
Request these from your platform team:
A Trino-capable compute pool.
The tenant-level
vulcan-trinostack (vulcan+trino:1.0) installed.One depot per source you read. The first depot must be a writable lakehouse depot; it's the materialization target.
Tenant secrets
vulcan-state-connection(Postgres) andvulcan-object-store-connection(S3), and a git-sync secret for your repo.Python 3.10 to 3.13 locally for the Vulcan CLI.
Confirm your depots are visible with dataos-ctl resource -t depot get -a.
Permissions
There are no Trino SQL grants to configure here: Vulcan connects to the internal coordinator over the Kubernetes network with method: no-auth. Access control lives entirely at each depot's underlying data source: whatever the depot's credentials permit there is what Trino can do through that catalog.
lakehouse
iceberg (REST)
DataOS-native; no external grants
bigquery
bigquery
Governed by IAM roles on the depot's service account
delta_lake
delta_lake
AWS IAM (S3 + Glue/Hive Metastore)
config.yaml
The gateway catalog is templated from an environment variable the stack injects, so the same file works locally and in production:
Cluster settings for this product live under deployment.trino: in this same file (see The deploy configuration below). Keep ignore_patterns: ["*-deploy.yaml"] so that a rendered deploy manifest saved inside the project is never parsed as a model.
Models: always use three-part names
Use fully-qualified catalog.schema.name for the model name and every source you read. The default catalog is the first secret catalog if any exist, otherwise the first depot. Unqualified names are a common cause of models materializing in the wrong place, so always spell out all 3 parts:
The deploy configuration
There's no separate manifest to hand-author for the Trino cluster. Coordinator and worker settings live under a deployment.trino: block, inside the same config.yaml you already have, alongside deployment.repo, deployment.compute, and the other keys every engine uses. See References → Vulcan → Deployment for the full deployment: walkthrough; this section covers only the Trino-specific parts.
The fields you'll change most often are deployment.depots (order matters: the first one is the default catalog and write target), deployment.trino.coordinator.resource, and deployment.trino.workers:
Size coordinator.resource and workers.resource to fit the compute pool; an undersized coordinator or worker limit is the most common deployment failure. TRINO_CATALOG must match a real catalog name.
Each deployment.trino.catalog.config entry is a "<tenant>:<secret-name>" reference to a DataOS secret whose key/value pairs are Trino catalog properties; the secret name becomes the catalog name. Secret catalogs take priority over depot catalogs as the default catalog, which is why the example pins TRINO_CATALOG to the lakehouse depot. Omit catalog if every source has a depot.
Set the JVM heap explicitly for both coordinator and workers, sized to the pod memory limit — an unmanaged heap is the most common cause of workers that OOM or never register. 4G is the validated -Xmx for a 4Gi pod limit; size yours to your own. If Postgres is the gateway catalog, set register_comments: false on the connection — the Postgres connector rejects the DDL Vulcan uses to register column comments.
Local development
Local development runs a real Trino coordinator and worker via Docker Compose (mounted at /usr/lib/trino/etc, Trino's real config path — a different mount path fails silently), plus a local Postgres container Trino federates against. Point config.yaml at the local coordinator (host: localhost, method: no-auth) with a DuckDB state_connection, add ignore_patterns: ["*-deploy.yaml"] so the deploy manifest is never parsed as a model, then validate with vulcan migrate && vulcan plan && vulcan run. See the Trino engine manual for the full Docker Compose setup.
Troubleshooting
Cluster forms with 0 workers / workers OOM at startup
Unmanaged JVM heap exceeds pod limit
Set -Xmx to fit the pod limit for both coordinator and workers.
Workers never register / cluster won't form
node.environment mismatch, or coordinator=true on a worker
Make node.environment identical on both roles; never set node.id on workers.
Custom plugin/UDF not loaded
JAR at dependencies/java/ root, not in a subdirectory
Move it into dependencies/java/<plugin>/.
Postgres catalog: vulcan plan fails on comment DDL
register_comments not disabled
Set register_comments: false on the gateway connection.
For the full failure-mode table, recovery procedures, and log locations, see Troubleshooting in the Trino engine manual.
Apply and verify
Deploying a dedicated Trino product uses the same single flow as every other engine: render the manifest from config.yaml and apply it; no separate dataos-ctl resource apply -f trino-deploy.yaml step is required.
Confirm the cluster is up by looking for SERVER STARTED in the coordinator logs. The Trino engine manual covers common issues (workers out of memory, wrong-catalog materialization, schedules expiring silently) and their fixes.
With the cluster running, continue to Assets.
Last updated
Was this helpful?