v1.90.2

版本发布时间: 2023-12-08 22:42:25

immich-app/immich最新发布版本:v1.119.1(2024-10-29 21:57:31)

v1.90.2

[!IMPORTANT]

Announcement (Breaking changes next release - v1.91.0)

Continuing the effort of reducing Immich's footprint, we would like to announce another planned change. Starting from the next release (not this release), we will be removing the Typesense container and changing the database image. Below are the changes that must be made in your docker-compose.yml file.

  immich-server:
  [...]
    depends_on:
      - redis
      - database
-     - typesense
    restart: always

  immich-microservices:
  [...]
    depends_on:
      - redis
      - database
-     - typesense
    restart: always

-  typesense:
-    container_name: immich_typesense
-    image: typesense/typesense:0.24.1@sha256:9bcff2b829f12074426ca044b56160ca9d777a0c488303469143dd9f8259d4dd
-    environment:
-      - TYPESENSE_API_KEY=${TYPESENSE_API_KEY}
-      - TYPESENSE_DATA_DIR=/data
-      # remove this to get debug messages
-      - GLOG_minloglevel=1
-    volumes:
-      - tsdata:/data
-    restart: always

[...]

  database:
    container_name: immich_postgres
-   image: postgres:14-alpine@sha256:6a0e35296341e676fe6bd8d236c72afffe2dfe3d7eb9c2405c0f3fc04500cd07
+   image: tensorchord/pgvecto-rs:pg14-v0.1.11
    env_file:
      - .env
    environment:

volumes:
  pgdata:
  model-cache:
- tsdata:

[!NOTE] Note: If you are running your database with a non-superuser role for Immich, you must enable the pgvecto.rs extension manually. You can do this by connecting to the immich database as a superuser and running:
CREATE EXTENSION vectors;

Search Background

It turns out that synchronizing data between Postgres and Typesense is quite complicated. It accounts for at least an additional 1000 lines of code and, while very feature-filled, has become a burden to maintain. Also, there have been some serious memory and performance issues with Typesense, especially with large photo collections. It is lightning fast, but at the expense of large indexes that must be read from disk into memory on each restart. There have been reports of this process taking up to 30 minutes!

We effectively remove an entire class of bugs and issues by removing the need to synchronize data. Furthermore, search queries can more easily be combined with the existing database schema, like libraries, partner sharing, albums, etc. After merging #3605, we hope to be able to make significant progress around the search implementation and feature set.

In short, some of the benefits of this change include the following:

One less container to run, deploy, and manage
Better memory/system resource usage
- Decreased memory usage
- Faster Encode CLIP and Recognize Faces jobs
- Faster startup time
Significantly less code to maintain, test, debug
Improved developer workflow when adding search-related functionality
Ability to integrate smart search into end-to-end testing for improved testing coverage and stability
Ability to integrate CLIP search with other (metadata) criteria in the future
Ability to add customizable index settings for even lower memory usage and higher performance, such as on-disk indices and quantization