gitlab-org--gitlab-foss/doc/development/uploads.md

Uploads development documentation

GitLab Workhorse has special rules for handling uploads. To prevent occupying a ruby process on I/O operations, we process the upload in workhorse, where is cheaper. This process can also directly upload to object storage.

The problem description

The following graph explains machine boundaries in a scalable GitLab installation. Without any workhorse optimization in place, we can expect incoming requests to follow the numbers on the arrows.

graph TB
    subgraph "load balancers"
      LB(HA Proxy)
    end

    subgraph "Shared storage"
       nfs(NFS)
    end

    subgraph "redis cluster"
       r(persisted redis)
    end
    LB-- 1 -->workhorse

    subgraph "web or API fleet"
      workhorse-- 2 -->rails
    end
    rails-- "3 (write files)" -->nfs
    rails-- "4 (schedule a job)" -->r

    subgraph sidekiq
      s(sidekiq)
    end
    s-- "5 (fetch a job)" -->r
    s-- "6 (read files)" -->nfs

We have three challenges here: performance, availability, and scalability.

Performance

Rails process are expensive in terms of both CPU and memory. Ruby global interpreter lock adds to cost too because the ruby process will spend time on I/O operations on step 3 causing incoming requests to pile up.

In order to improve this, disk buffered upload was implemented. With this, Rails no longer deals with writing uploaded files to disk.

graph TB
    subgraph "load balancers"
      LB(HA Proxy)
    end

    subgraph "Shared storage"
       nfs(NFS)
    end

    subgraph "redis cluster"
       r(persisted redis)
    end
    LB-- 1 -->workhorse

    subgraph "web or API fleet"
      workhorse-- "3 (without files)" -->rails
    end
    workhorse -- "2 (write files)" -->nfs
    rails-- "4 (schedule a job)" -->r

    subgraph sidekiq
      s(sidekiq)
    end
    s-- "5 (fetch a job)" -->r
    s-- "6 (read files)" -->nfs

Availability

There's also an availability problem in this setup, NFS is a single point of failure.

To address this problem an HA object storage can be used and it's supported by direct upload

Scalability

Scaling NFS is outside of our support scope, and NFS is not a part of cloud native installations.

All features that require Sidekiq and do not use direct upload won't work without NFS. In Kubernetes, machine boundaries translate to PODs, and in this case the uploaded file will be written into the POD private disk. Since Sidekiq POD cannot reach into other pods, the operation will fail to read it.

How to select the proper level of acceleration?

Selecting the proper acceleration is a tradeoff between speed of development and operational costs.

We can identify three major use-cases for an upload:

1. storage: if we are uploading for storing a file (i.e. artifacts, packages, discussion attachments). In this case direct upload is the proper level as it's the less resource-intensive operation. Additional information can be found on File Storage in GitLab.
2. in-controller/synchronous processing: if we allow processing small files synchronously, using disk buffered upload may speed up development.
3. Sidekiq/asynchronous processing: Async processing must implement direct upload, the reason being that it's the only way to support Cloud Native deployments without a shared NFS.

For more details about currently broken feature see epic &1802.

Handling repository uploads

Some features involves Git repository uploads without using a regular Git client. Some examples are uploading a repository file from the web interface and design management.

Those uploads requires the rails controller to act as a Git client in lieu of the user. Those operation falls into in-controller/synchronous processing category, but we have no warranties on the file size.

In case of a LFS upload, the file pointer is committed synchronously, but file upload to object storage is performed asynchronously with Sidekiq.

Upload encodings

By upload encoding we mean how the file is included within the incoming request.

We have three kinds of file encoding in our uploads:

1. multipart: multipart/form-data is the most common, a file is encoded as a part of a multipart encoded request.
2. body: some APIs uploads files as the whole request body.
3. JSON: some JSON API uploads files as base64 encoded strings. This will require a change to GitLab Workhorse, which is planned.

Uploading technologies

By uploading technologies we mean how all the involved services interact with each other.

GitLab supports 3 kinds of uploading technologies, here follows a brief description with a sequence diagram for each one. Diagrams are not meant to be exhaustive.

Rack Multipart upload

This is the default kind of upload, and it's most expensive in terms of resources.

In this case, workhorse is unaware of files being uploaded and acts as a regular proxy.

When a multipart request reaches the rails application, Rack::Multipart leaves behind tempfiles in /tmp and uses valuable Ruby process time to copy files around.

sequenceDiagram
    participant c as Client
    participant w as Workhorse
    participant r as Rails

    activate c
    c ->>+w: POST /some/url/upload
    w->>+r:  POST /some/url/upload

    r->>r: save the incoming file on /tmp
    r->>r: read the file for processing

    r-->>-c: request result
    deactivate c
    deactivate w

Disk buffered upload

This kind of upload avoids wasting resources caused by handling upload writes to /tmp in rails.

This optimization is not active by default on REST API requests.

When enabled, Workhorse looks for files in multipart MIME requests, uploading any it finds to a temporary file on shared storage. The MIME data in the request is replaced with the path to the corresponding file before it is forwarded to Rails.

To prevent abuse of this feature, Workhorse signs the modified request with a special header, stating which entries it modified. Rails will ignore any unsigned path entries.

sequenceDiagram
    participant c as Client
    participant w as Workhorse
    participant r as Rails
    participant s as NFS

    activate c
    c ->>+w: POST /some/url/upload

    w->>+s: save the incoming file on a temporary location
    s-->>-w:  

    w->>+r:  POST /some/url/upload
    Note over w,r: file was replaced with its location<br>and other metadata

    opt requires async processing
      r->>+redis: schedule a job
      redis-->>-r:  
    end

    r-->>-c: request result
    deactivate c
    w->>-w: cleanup

    opt requires async processing
      activate sidekiq
      sidekiq->>+redis: fetch a job
      redis-->>-sidekiq: job

      sidekiq->>+s: read file
      s-->>-sidekiq: file

      sidekiq->>sidekiq: process file

      deactivate sidekiq
    end

Direct upload

This is the more advanced acceleration technique we have in place.

Workhorse asks rails for temporary pre-signed object storage URLs and directly uploads to object storage.

In this setup an extra rails route needs to be implemented in order to handle authorization, you can see an example of this in Projects::LfsStorageController and its routes.

note: this will fallback to disk buffered upload when direct_upload is disabled inside the object storage setting. The answer to the /authorize call will only contain a file system path.

sequenceDiagram
    participant c as Client
    participant w as Workhorse
    participant r as Rails
    participant os as Object Storage

    activate c
    c ->>+w: POST /some/url/upload

    w ->>+r: POST /some/url/upload/authorize
    Note over w,r: this request has an empty body
    r-->>-w: presigned OS URL

    w->>+os: PUT file
    Note over w,os: file is stored on a temporary location. Rails select the destination
    os-->>-w:  

    w->>+r:  POST /some/url/upload
    Note over w,r: file was replaced with its location<br>and other metadata

    r->>+os: move object to final destination
    os-->>-r:  

    opt requires async processing
      r->>+redis: schedule a job
      redis-->>-r:  
    end

    r-->>-c: request result
    deactivate c
    w->>-w: cleanup

    opt requires async processing
      activate sidekiq
      sidekiq->>+redis: fetch a job
      redis-->>-sidekiq: job

      sidekiq->>+os: get object
      os-->>-sidekiq: file

      sidekiq->>sidekiq: process file

      deactivate sidekiq
    end