Trino Experiments

HEAVILY OUT-OF-DATE: will be cleaned up in the coming days

---

This repository contains a small Trino experiment stack:

• doe-suite-config/ defines the DoE-Suite experiment that provisions and starts Trino.
• queries/ contains the TPC-H SQL templates used by the loader.
• trino-loader/ submits TPC-H queries to Trino and records Trino query stats.
• plots/plot_trino_split_spans.py extracts split spans from Jaeger and writes SVG/CSV analysis outputs.

Prerequisites

Install the local tools used by this repo:

• Python >=3.9,<3.11, matching doe-suite-config/pyproject.toml.
• Poetry, for doe-suite-config.
• Rust and Cargo, for trino-loader.
• SSH access to the machines listed in the selected DoE-Suite inventory.
• SSH agent forwarding so remote machines can clone this private Git repository.

Initialize the DoE-Suite submodule after cloning:

git submodule update --init --recursive

Set the DoE-Suite project variables from the repository root:

export DOES_PROJECT_DIR="$PWD"
export DOES_PROJECT_ID_SUFFIX="<your-short-id>"

doe-suite-config/group_vars/all/main.yml builds the remote project id as dataprocessing_<DOES_PROJECT_ID_SUFFIX> and uses /tmp/<project-id> as the remote experiment folder.

Set Up a Trino Cluster With DoE-Suite

The Trino suite is defined in doe-suite-config/designs/trino.yml. Set host_types.worker.n there to the number of Trino workers you want to use. The selected inventory must contain at least that many machines in the worker group. By default, the design creates one master host and one worker host.

Create or edit an inventory before running:

• doe-suite-config/inventory/cloudlab.yml has master_01 and worker_01 entries with CloudLab hostnames and private IPs.

The Trino configuration templates are in doe-suite-config/roles/trino/tasks/: config.properties, node.properties, jvm.config, and catalog/s3data.properties. The role renders these into the remote Trino etc/ directory on both master and worker hosts.

Configure Trino's filesystem cache in doe-suite-config/roles/trino/tasks/catalog/s3data.properties. The relevant settings are fs.cache.enabled, fs.cache.directories, and fs.cache.max-sizes; switch the cache directory between /tmp/trino-cache and /dev/shm/trino-cache depending on whether you want disk-backed or memory-backed cache.

The same s3data.properties file also contains the S3 and Glue catalog settings. Add the AWS S3 access key and secret key there before running against private TPC-H data.

The setup-master role also installs Prometheus, Grafana, and Jaeger. setup-base installs prometheus-node-exporter on every host. Trino OpenTelemetry tracing is enabled by default in doe-suite-config/group_vars/all/main.yml and points to Jaeger's OTLP gRPC endpoint on the master.

Run the suite from the doe-suite submodule, selecting an inventory by filename without .yml:

cd doe-suite
make run suite=trino id=new cloud=cloudlab

After setup, connect to the master services directly or through SSH port forwarding:

ssh -L 8080:localhost:8080 -L 9090:localhost:9090 -L 16686:localhost:16686 -L 3000:localhost:3000 <master-host>

With those forwards open, use these local service URLs:

• Trino coordinator: http://localhost:8080
• Prometheus: http://localhost:9090
• Jaeger: http://localhost:16686
• Grafana: http://localhost:3000

Submit TPC-H Queries With trino-loader

trino-loader expects the TPC-H data to be available in Trino as:

s3data.tpch_sf<SCALE_FACTOR>.<table>

The SQL files in queries/q1.sql through queries/q22.sql use templates like `DATASET.lineitem`. At runtime, trino-loader rewrites those references to s3data.tpch_sf<SCALE_FACTOR>.lineitem, strips the trailing semicolon, posts the statement to /v1/statement, follows nextUri until completion, and records the final Trino stats JSON.

Build the loader:

cd trino-loader
cargo build --release
cd ..

Run one query:

./trino-loader/target/release/trino-loader \
  --master <master-host-or-ip> \
  --port 8080 \
  --scale-factor 1 \
  --template-path queries \
  --query-number 1

Analyze Split Spans

plots/plot_trino_split_spans.py is a standard-library Python script. It can either fetch live data from Trino, Jaeger, and Prometheus, or read saved JSON files.

With the SSH port forwards above open, run:

python3 plots/plot_trino_split_spans.py

The script defaults to Trino on localhost:8080, Jaeger on localhost:16686, and Prometheus on localhost:9090. It selects the latest traced non-failed Trino query unless --query-id is given. Outputs are written as trino_split_spans_<queryId>.svg, trino_split_spans_<queryId>.csv, trino_child_spans_<queryId>.csv, and trino_stage_graph_<queryId>.svg.

To analyze a downloaded Jaeger trace instead of fetching it live:

python3 plots/plot_trino_split_spans.py --trace-json <trace.json>