Convert AWS S3 Access Logs to Parquet

This repo provides a sample pipeline using Clusterless and Tessellate to convert native AWS S3 access logs into Apache Parquet files. Additionally, a Glue database and table will be created to allow for SQL queries against the data.

See https://github.com/ClusterlessHQ/clusterless-aws-examples for simpler examples that should be run first to familiarize yourself with Clusterless.

The code in this repo is free to fork, copy, and use under the MIT License.

Overview

This example expects that an existing bucket in your account has server access logging enabled. The location of the access logs will need to be added to the .properties.libsonnet file (see below).

The project file, aws-logs-arc-project.jsonnet, when deployed, will create five core resources.

• A new AWS S3 bucket to store the result dataset as Parquet files.
• A Clusterless Boundary (a Clusterless provided AWS Lambda), to listen to the logging location for new log files in order to push availability events.
• A Clusterless Arc to listen for new log availability events and convert the text files into Parquet.
• An AWS Glue database and an AWS Athena queryable table.
• And a second arc workload that will add new partitions to the new table as they arrive.

The first arc workload is simply a Tessellate instance running as a Docker image in AWS Fargate. When the arc completes, a new availability event is published allowing for other arcs to participate in the pipeline.

Every time AWS drops a new log file in the access logging bucket, the boundary is notified by AWS. The boundary in turn, every 5 minutes, creates a manifest of the newly arrived file(s) and publishes an availability event for the access log dataset.

The arc listens for new arrivals of data from the access log dataset and runs the Tessellate workload. On completion, a new dataset availability event is published.

If no data arrives in the access logging bucket, an event is still published every 5 minutes, but the manifest is marked as empty. This way it's clear the system is running, but there just happens to be no new data. The Tessellate arc still runs, but it in turn also publishes an empty manifest (a future version will short-circuit the workload and publish an empty manifest to reduce costs).

Notes

This example relies on jsonnet to generate the JSON files that are piped into the cls command via cls -p -.

For this to work, --approve must be included in the cls command, which may seem a bit scary.

Alternatively, pipe the jsonnet output to a file and then run cls on the file without --approve to enable manual approvals.

By editing app/ingest.json, this sample can trivially parse other log formats into Parquet. See the Tessellate documentation for more information.

Running

Copy .properties.libsonnet.template to .properties.libsonnet and edit.

Tessellate should also be installed locally to test the arc locally. Otherwise, this step may be skipped.

Note that .properties.libsonnet is listed in the .gitignorefile. If forking this project to use as a template for an actual deployment, remove the declaration from the ignore file, so it can be maintained in your git repo.

Bootstrap

The following command will initialize the AWS placement environment, if not already done so.

jsonnet bootstrap.jsonnet | jq -r '.[0]' | sh

Verify

The following command will confirm the jsonnet file translates into a valid clusterless project file.

jsonnet aws-logs-arc-project.jsonnet | cls verify -p -

Deploy the Boundary

These next two steps are optional, you may skip directly to Deploy to simply deploy the whole project.

The provided project file deploys both a boundary and an arc. We need to deploy the boundary first so that we can locally test the arc.

jsonnet aws-logs-arc-project.jsonnet | cls deploy -p - --exclude-all-arcs --approve

Because we are piping the project file into cls, we must include the --approve option.

Note, this pipeline could have been split into multiple projects files, one with resources and boundaries, and another with arcs. For this example we are using one file.

To show the status of the dataset created by the boundary, call:

cls datasets status

You should expect to see the summary for s3-access-logs.

Alternatively, log into S3 to confirm new manifests are being created. They will be located in:

s3://[STAGE]-clusterless-manifest-[ACCOUNT]-[REGION]/datasets/name=s3-access-logs/

If manifests do not show up in a few minutes, confirm logs are being written to the given log path specified in the .properties.libsonnet file.

Once confirmed, check to see if the deployed boundary lambda is running successfully in the AWS Console.

Test Locally

Once manifests begin to arrive, copy a lot id that has a state=complete state and use it below:

jsonnet aws-logs-arc-project.jsonnet | cls local -p - --arc convertParquet --lot 20230717PT5M250 > local.sh

Run local.sh from the app directory:

chmod a+x local.sh
cd app
../local.sh

Currently, this requires Tessellate to be installed locally. A future version will build and execute the Docker image (which automatically downloads and install tess into the image).

It's encouraged to look inside the local.sh script that was generated. The shell script emulates how the resulting Docker image will be executed in AWS Batch.

To test an empty manifest, look for a lot with state=empty, if any.

jsonnet aws-logs-arc-project.jsonnet | cls local -p - --arc convertParquet --lot 20230731PT5M217 --manifest-state empty > local.sh

If local.sh completes with no errors, confirm a manifest file, for the given lot, was written to:

s3://[STAGE]-clusterless-manifest-[ACCOUNT]-[REGION]/datasets/name=access-logs-parquet/

Look inside the manifest file to see where the data was written and confirm it is available.

Note that all the Clusterless metadata is human-readable as JSON text files. And much of the important information is embedded in the file path and name (AWS S3 object key), for example, the state of the manifest file (empty or complete), or the state of the arc itself. So reporting on running arcs is as simple as a aws s3 ls ... command.

Deploy

This command will deploy the declared project.

jsonnet aws-logs-arc-project.jsonnet | cls deploy -p - --approve

Actually, the whole project will be deployed, the resources and boundaries deployed above will be "updated" (if previously deployed as above), and the arc will be created.

Note that if changes are made to the project file, usually it can just be re-deployed for the updates to take effect. This allows a developer to incrementally deploy an updated arc workload for development when testing locally isn't possible.

Now to confirm the deployment.

To show the project was deployed.

cls projects

To show the dataset status.

cls datasets status

When everything is running, there will be three summaries.

Finally, to show the arc status.

cls arcs status

In time, you will see two summaries.

Destroy

This command will destroy the whole project and clean up what resources it can.

jsonnet aws-logs-arc-project.jsonnet | cls destroy -p - --approve

By default, Clusterless does not destroy most resources like AWS S3 Buckets or AWS Glue Databases and Tables. They will need to be removed manually.

This can be overridden via local/global configuration values.

cls config init

Then edit the local .clsconfig file.

To print the current configuration:

cls config show