# GS Caltex links no-code modeling to scheduled batch inference

- Source: MechNext — https://mechnext.eu/ru/article/71-gs-caltex-links-no-code-modeling-to-scheduled-batch-inference
- Published: 2026-11-14
- Updated: 2026-11-14
- Category: Industry News
- Language: ru (Markdown in other languages: ?lang=en, ?lang=de, ?lang=sk, ?lang=pl, ?lang=zh)

> GS Caltex describes an automated handoff from SageMaker Canvas to SageMaker Pipelines that reuses existing assets, reduces deployment from about a week to a few hours, and runs batch inference every 30 minutes.

According to the AWS case study, GS Caltex divides model work between process engineers and ML engineers. Process engineers build and validate models without writing code in Amazon SageMaker Canvas. ML engineers then configure Amazon SageMaker Pipelines so that validated work can be deployed and used for scheduled inference.

The handoff between these roles is automated and reuses assets created in Canvas. It does not remove the production configuration performed by ML engineers. They still set production data paths, model settings, preprocessing, and inference jobs. The arrangement therefore carries validated assets forward while retaining an explicit engineering stage for production execution.

### From a validated model to a production resource

The deployment process registers a validated Canvas model version in Model Registry, checks whether the model package is approved, and creates a SageMaker model resource that Batch Transform can reference. Model Registry is also used to manage model versions within the reference architecture.

The implementation reuses a Data Wrangler preprocessing flow as well. It exports that flow through SageMaker Processing and configures a pipeline step with the resulting processing-job settings and the associated `.flow` file. This connects the preprocessing flow already used during model work with the scheduled production workflow.

### A five-step inference workflow

The sample inference workflow contains five defined steps:

- prepare the input,
- apply preprocessing,
- remove the target column,
- run batch inference,
- assemble the results.

For each scheduled execution, the workflow prepares input data from Amazon S3, performs preprocessing, runs SageMaker Batch Transform, and stores the resulting output in S3. Operators can inspect individual step status and Amazon CloudWatch logs for monitoring.

A specific data-shape issue must be handled because inference data does not contain the training target column. In the reference implementation, a placeholder target with a compatible type and position is temporarily added before preprocessing and removed afterward. The method is not suitable without additional treatment if the target influences feature values or transformations. That limitation matters when deciding whether the same procedure can be used for a particular preprocessing flow.

### Scheduling and its operational boundary

Amazon EventBridge can start the pipeline with a rate expression every 30 minutes. This schedule produces 48 planned runs per day. A scheduled start does not, however, guarantee that one execution will finish before the next begins. Runtime and the possibility of overlapping executions therefore remain conditions that must be evaluated for production use.

GS Caltex reports that model deployment now takes a few hours rather than about a week. The company also runs batch inference every 30 minutes and reports that avoiding continuously running endpoints has meaningfully reduced inference costs. These are results reported by GS Caltex; the supplied source does not establish them as independently verified findings.

### Checks required before production scheduling

Before recurring production runs are enabled, the workflow should be validated for schema and output consistency. Validation should also confirm the association between each source record and its prediction, so that the assembled output remains connected to the corresponding input.

Testing should cover failure behavior, total runtime, and cost. It should also address overlapping executions and the possibility that the same input window could be processed more than once. A 30-minute rate defines when executions are started; it neither proves completion within 30 minutes nor prevents duplicate processing by itself.

The reference architecture consequently combines no-code model creation, managed model versions, reused preprocessing, batch execution, S3-based input and output, and EventBridge scheduling. Its production operation still depends on the ML engineers' configuration of data paths and jobs, approval of the registered model package, and validation of behavior under the intended schedule. The reported reduction in deployment time illustrates the value GS Caltex attributes to automating the handoff, while the validation requirements define the operational work that remains.
