Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarAWS Machine Learning Blog→ original

GS Caltex Cuts ML Model Deployment from a Week to a Few Hours

Models built by domain experts often wait weeks for engineers to move them into production. GS Caltex automated that handoff from AWS SageMaker Canvas, a no-code tool. Deploying a model now takes a few hours instead of roughly a week. Predictions run in batches every 30 minutes, so the server does not need to stay on around the clock. If your models are built by technologists rather than programmers, this kind of bridge can save engineers days of work.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
GS Caltex Cuts ML Model Deployment from a Week to a Few Hours
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Five stepspreparation, preprocessing, prediction, result assembly
Key metric #2EventBridge triggers the job 48 times a day, once every 30 minutes
Check whether runs overlapfinishing within 30 minutes is not guaranteed

Imagine a plant where a process engineer spends months tuning process parameters and finally gets a model that predicts the target metric well. Next comes the dullest part: turning the model into a job that runs on a schedule against fresh data. This stage usually takes longer than training. South Korean energy chemicals company GS Caltex described in an AWS blog post how it cut that stage from roughly a week to a few hours.

A quick breakdown. An ML model here is a trained forecasting program. Deployment to production means moving it from the experimental environment into the live system. SageMaker Canvas is AWS's visual tool for building models without code. SageMaker Pipelines is a service that links data preprocessing and prediction steps into an automated pipeline. The company connected these two tools.

Who Does What

GS Caltex split the roles according to each group's strengths. Process engineers work in Canvas: they prepare data with the Data Wrangler visual editor, train the model, and assess how far the predictions can be trusted. They know the process, the variables, and the real data. ML engineers take the finished results and turn them into production jobs: they manage model versions, run history, and failure handling.

The bottleneck was the handoff. Moving assets from Canvas to Pipelines had to be done by hand: writing data paths for production, passing the model configuration, and wiring the preprocessing and prediction steps together. GS Caltex automated that handoff and reuses everything the engineer has already validated. According to the company, ML engineers now deploy models in a few hours. Previously it took about a week.

How the Pipeline Works

The scheme has two parts. The first runs when a new model version is released. An engineer receives the preprocessing file and the model from Canvas, checks data formats and prediction results, approves the model in the Model Registry, and deploys it. The second part runs continuously: Amazon EventBridge triggers a five-step pipeline on a schedule.

  • PrepareInput adds a placeholder target-metric column to the data if it is missing;
  • PreprocessNewData runs the data through the same preprocessing the engineer validated;
  • DropTargetColumn removes the placeholder and the CSV header;
  • BatchPredict runs batch prediction with the approved model;
  • AssembleResults builds the final CSV from the prepared data and predictions.

The placeholder is needed for a simple reason. The processing flow was built on training data, where the known outcome exists. New data does not have it by definition, yet the flow still expects that column. The authors add a caveat: the approach works only if the placeholder does not affect feature values. If the flow filters rows by the target metric, builds features from it, or applies target encoding, it needs separate rework.

Prediction runs in batches every 30 minutes, which means 48 runs a day. Between runs, there is no need to keep a constantly running endpoint, meaning a server that waits for requests. The company says this noticeably reduced inference costs, but it does not provide specific figures. For tasks that require an instant response, batch mode will not work.

Pitfalls

The authors candidly list the places where you can trip up:

  • A fixed input data path. The PrepareInput step writes data to a location specified inside the .flow file. A pipeline parameter cannot replace that path on its own. To isolate each run, you need a step that updates the path in a copy of the file. Runs that share an input path must not overlap.
  • Model approval is not checked automatically. The Model Registry stores the status but does not block predictions. The check has to be built into the deployment code, and schedules must be updated if an approval is revoked.
  • Learned parameters. Scaling and encoding fitted during training should not be recomputed on every new batch. It is better to pin the container image version.
  • Consistency checks. Compare columns, data types, and values after processing in Canvas and in production, using the same input.

What to Take Away

The lesson reaches beyond a single AWS stack. If domain experts build the models, the value comes from a standard bridge between their tool and production, not from yet another tool. A reference implementation is on GitHub and is built around a credit classification model example, so the configuration will need to be adjusted to your input and output formats. When assessing the savings, account for the whole chain, not just the server's hourly rate: preprocessing, instance startup, prediction, S3 storage, and logs.

Story

Key Questions Answered

Who Does What?

GS Caltex split the roles according to each group's strengths. Process engineers work in Canvas: they prepare data with the Data Wrangler visual editor, train the model, and assess how far the predictions can be trusted. They know the process, the variables, and the real data. ML engineers take the finished results and turn them into production jobs: they manage model versions, run history, and failure handling. The bottleneck was the handoff.

How the Pipeline Works?

The scheme has two parts. The first runs when a new model version is released. An engineer receives the preprocessing file and the model from Canvas, checks data formats and prediction results, approves the model in the Model Registry, and deploys it. The second part runs continuously: Amazon EventBridge triggers a five-step pipeline on a schedule.

What are the key details on Pitfalls?

The authors candidly list the places where you can trip up: A fixed input data path. The PrepareInput step writes data to a location specified inside the .flow file. A pipeline parameter cannot replace that path on its own. To isolate each run, you need a step that updates the path in a copy of the file. Runs that share an input path must not overlap. Model approval is not checked automatically. The Model Registry stores the status but does not block predictions.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…