Weblate internals

Napomena

This chapter will give you basic overview of Weblate internals.

Weblate derives most of its code structure from, and is based on Django.

Directory structure

Quick overview of directory structure of Weblate main repository:

docs

Source code for this documentation, which can be built using Sphinx.

dev-docker

Docker code to run development server, see Running Weblate locally in Docker.

weblate

Source code of Weblate as a Django application, see Weblate internals.

weblate/static

Client files (CSS, Javascript and images), see Weblate sučelje.

Moduli

Weblate consists of several Django applications (some optional, see Optional Weblate modules):

accounts

Korisnički račun, profili i obavijesti.

addons

Dodatci za podešavanje Weblatea, pogledaj Dodaci.

api

API based on Django REST framework.

auth

Autentifikacija i dozvole.

billing

Opcionalni modul Naplata.

checks

Translation string Provjere kvalitete module.

fonts

Font rendering checks module.

formats

File format abstraction layer based on translate-toolkit.

gitexport

The optional Git exporter module.

lang

Module defining language and plural models.

legal

The optional Legal module module.

machinery

Integracija usluga strojnog prevođenja.

memory

Ugrađena prevodilačka memorija, pogledaj Prevodilačka memorija.

screenshots

Screenshots management and OCR module.

trans

Main module handling translations.

utils

Various helper utilities.

vcs

Version control system abstraction.

wladmin

Django admin interface customization.

Translation sources

A unit has a canonical file source and may have a translation from another language as its effective source. Keep Unit.source and Unit.source_unit canonical: repository synchronization and file formats depend on this. Use effective_source and its language and plural accessors for content shown to translators. Do not replace source on a copied unit to simulate an effective source; other fields and related objects still describe the canonical source.

Use Unit.source_snapshot when an operation needs to preserve the source text, language, plural rules, and terminology metadata together. Its identity defines source changes for dependency reconciliation, editor conflict detection, and AI diagnostics. For historical rendering and reports, read the snapshot recorded on the change, rather than reconstructing it from the current unit.

A missing configured parent remains a custom source: its text is empty, its language is retained, and its child is blocked. effective_source_unit is then None. Only removing the workflow or selecting the canonical language restores the canonical source.

Pending source snapshots retain the first source change until acceptance, including for targets already needing editing. Their previous_state metadata prevents source reversion from accepting a target that already needed editing. Without a previous snapshot, custom sources do not use the canonical previous_source.

Text-only accessors do not load language or plural metadata. For batch consumers that need the complete effective source, use UnitQuerySet.prefetch_translation_parent() alongside their existing unit prefetches. prefetch_source() already includes this. Ordinary searches keep the canonical query path and ordinary units retain their existing content hashes.

Source-dependent writes enter source_operation() before locking unit rows. The outer operation takes a shared project gate and reads its workflow mapping in one database round trip, including for projects without custom workflows. Workflow changes take source_project_gate(..., exclusive=True) before validation or persistence. Configuration changes affecting several projects acquire their gates in ascending project ID order. Never upgrade a shared gate to an exclusive one inside an operation.

The database gate function locks first and reads workflows in a separate statement, so a waiter sees configuration committed while it waited. Write paths use this transaction-scoped view rather than cached model settings. Nested operations reuse it; rolled-back savepoints discard their dependency work. Queued dependencies are reconciled before commit and external memory updates are scheduled after commit. Source-change history is written in bulk, preserving individual records while batching their notification dispatch.

Workflow saves and deletions compare the old and new source mappings under the exclusive gate. Reconciliation visits changed languages and their transitive dependants in either mapping, skipping components without affected translations. It loads supporting parent units without rebuilding unrelated languages. Callers that need a complete rebuild, such as restoration and language maintenance, omit the language restriction.

Within each component, reconciliation updates relationships before running checks. Workflow changes run local checks with the existing component batching enabled, then schedule expensive checks through the component check finalizer after commit. Boundary-wide batches also clear stale warnings on old-source peers. Successful child batches are combined at the exclusive project gate, deduplicating shared project and workspace checks while preserving rollback boundaries and actor attribution. Eager tasks and automation contexts finalize their batches inline. Ordinary dependency reconciliation retains its synchronous unit and previous-source peer checks.

Translation-memory payloads use bounded batches scheduled after commit. Nested operations merge dependency work, while rolled-back savepoints discard their work and commit callbacks.

Acquire any required repository locks before project gates, then component and unit row locks. Reconciliation itself only changes database state and does not acquire repository locks. Apply this boundary to bulk imports, automatic translation, file-save failures, deletion, and restoration as well as editor writes. Language and plural-rule changes require exclusive gates before their updates. Historical migrations use historical models without these runtime helpers.

Machinery source selection happens before mapping language codes to a provider. Use get_unit_source_language() for translation and uses_custom_source() for workflow-specific glossary handling. Provider codes can merge distinct language variants, so code equality cannot determine whether a workflow source was selected. Explicit component or secondary source selection takes precedence over the workflow source.

Background tasks internals

Savjet

This section describes Celery task internals. Background tasks using Celery describes how to configure Celery to run the tasks.

Weblate uses Celery to execute tasks in the background. Some tasks are event-triggered, and some tasks are schedule-triggered.

The Celery Beat is used for scheduling tasks, and django-celery-beat is used to store the periodic task schedule in the database. The tasks schedule is configured in tasks.py in each of the Django apps.

The tasks are consumed using several queues; the routing is configured in settings.py. The queues were designed to separate different types of workload:

celery

The default queue where background tasks are processed.

notify

Delivers notification e-mails, both for events within Weblate and for authentication or registration. This is a separate queue to make e-mail delivery smooth even if there is a backlog of other tasks.

memory

Updates translation memory entries. The updating queue can be long when importing new strings, and long processing does not matter much here, so having a separate queue avoids blocking other tasks.

backup

The backup tasks cannot be executed in parallel, and a single dedicated worker makes this easier.

translate

Automatic translation tasks are known to take long because they hit external services.