• Title: Auto-Batching

# Auto-Batching

# 1. Summary

Meilisearch can automatically group consecutive asynchronous documentAddition or documentPartial tasks for the same index via an automatic batching mechanism.

The user can disable this auto-batching behavior. See 3.2. Auto-batching mechanisms options section.

# 2. Motivation

We have regularly collected user pain points pointing out the slow indexing over the last year. We explained several times to users to make batches containing a maximum of documents to be updated/added to compress the indexing time of specific data structures.

To make Meilisearch easier to use, we explored the idea of automatically creating these batches within Meilisearch before indexing users’ documents.

# 3. Functional Specification

# 3.1. Explanations

A batch preserves the logical order of the tasks for a given index.

Only consecutive documentAdditionOrUpdate tasks for the same index can be in the same batch. All tasks concerning other operations will also be part of a batch having only one task.

# 3.1.1. Grouping tasks to a single batch

The scheduling program that groups tasks within a single batch is triggered when an asynchronous task currently processed reaches a terminal state as succeeded or failed.

In other words, when a scheduled documentAdditionOrUpdate task for a given index is picked from the task queue, the scheduler fetches and groups all documentAdditionOrUpdate tasks for that same index in a batch.

The more similar consecutive tasks the user sends in a row, the more likely the batching mechanism can group these tasks.

# Schema

Auto-batching Process

# batchUid generation

All tasks are part of a batch identified by an internal batchUid field. A task batch preserves the logical order of the tasks for a given index. The batch identifiers are unique and strictly increasing. The batchUid field is internal; thus not visible on a task resource.

# 3.1.2. Impacts on task API resource

  • The different tasks grouped in a batch are processed within the same transaction. But if a task fails within a batch, the whole batch does not fail, only the related task.
  • Tasks within the same batch share the same values for the startedAt, finishedAt, duration fields, and the same error object if an error occurs for a task during the batch processing.
  • If a batch contains many tasks, the task details indexedDocuments is identical in all tasks belonging to the same processed batch.

# 4. Technical Aspects


# 5. Future Possibilities

  • Extends it for all consecutive payload types.
  • Expose the batchUid field and add a filter capability on it on the /tasks endpoints.
  • Report the documents that could not be indexed to the user in a more precise manner.
  • Optimize some tasks sequence, for example if there is a document addition followed by an index deletion, we could skip the document addition.