dali (named after the great painter Salvador Dali) is a service which performs image transformations. It is used by OLX to process and serve images for users in around 40 countries. The application supports:
- Retrieving source images from a HTTP URL
- Encoding images to PNG, JPEG, WEBP or HEIC
- Resizing an image
- Rotating an image
- Apply a watermark image to an image
All configuration should be provided through either a json config file or environment variables. The application will search first by a file named config/default.json, then if the environment variable RUN_MODE is set it will search for config/RUN_MODE, override the values and lastly it will override with environment variables.
| Name | Type | Description | Required | Possible Values | Notes |
|---|---|---|---|---|---|
log_level |
Enum(trace, debug, info, warn, error) | Logging level for the application | N |
|
Default value is info. |
app_port |
integer | Port which the web server listens to for requests | Y | - | |
health_port |
integer | Port which the web server listens to for the health requests | Y | - | |
server_client_timeout |
integer | Defines a timeout for reading client request header. If a client does not transmit the entire set headers within this time, the request is terminated with the 408 (Request Time-out) error. | N | - | - |
client_shutdown_timeout |
integer | Defines a timeout for shutdown connection. If a shutdown procedure does not complete within this time, the request is dropped. | N | - | - |
server_keep_alive |
integer | Server keep alive value | N | - | If not specified, it will use the OS's configuration |
http_client_con_timeout |
integer | Set the timeout for connecting to a URL | N | - | Default is no timeout |
http_client_read_timeout |
integer | Set the timeout for the response | N | - | Default is no timeout |
http_client_write_timeout |
integer | Set the timeout for the request | N | - | Default is no timeout |
http_client_max_size_of_payload |
integer | Max size of response payload (for awc feature only) |
N | - | Default is 256Kb |
max_threads |
integer | Max number of threads the application will spawn for serving requests and processing images | N | - | If not specified it will take the number of physical CPUs from the machine |
vips_threads |
integer | Max number of threads for image processing that will be used | N | - | if not specified it will take max_threads/2 with a minimum of 1 |
app_threads |
integer | Max number of threads for serving requests that will be used | N | - | if not specified it will take max_threads/2 with a minimum of 1 |
metrics_threads |
integer | Max number of threads for serving the health and metrics endpoints |
N | - | Defaults to 1 |
The application will compute the number of threads by the following formula: pod_number_of_cpus * cpu_usage_percentage / 100. This number will be divided by 2 and half will be assigned to the HTTP connection listener and half will be assigned to libvips (the image library). An extra worker will be created to listen to the health endpoint (this was done to be sure the application won't block the health endpoint even when overloaded).
- Libvips
- A HTTP server for images
- Docker
- Rust
This application relies on C libvips library. That means it has to be previously installed into the system before compiling and/or running.
For installation follow this instructions. (Required minimum version 8.8.3)
Using rustup is the recommended way to install rust. It is a tool that manages and updates rust versions (like nvm for node for example). To install it, simply run curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh. Then run rustup update.
To build and run the application, run the following command:
cargo run
Note: if you're building the application on a linux musl system, you'll need to set this env variable: RUSTFLAGS='-C target-feature=-crt-static'
if it fails to compile because of the linker, try to add the result of pkg-config --libs vips to RUSTFLAGS= env variable`
Alternatively, it is possible to run both components inside docker (once you build them):
make up
There are 3 kinds of tests: unit, integration and benchmark tests.
At the moment, the only logic the application has is regarding resizing and watermark positioning. There are some unit tests around these use cases in the package image_processor. All other functions are deeply tied to the libvips library which complicates testing since it is not easily mockable.
To run them, simply execute: cargo test --bin dali. The parameter --bin is needed because otherwise it would run also integration tests.
This tests run over a running application.
To run the tests, the script will start the containers through docker-compose, copy some sample files to the http container and run the tests over the application, checking the array of bytes from the responses against expected result images stored in the tests/resources/results directory. To run the whole flow, simply run: make test.
This is an experimental feature from rust, so in order to use, the nightly rust toolchain has to be enabled. To do that run:
rustup toolchain install nightly
rustup default nightly
To rollback to stable, run rustup default stable.
To run the benchmark, simply call cargo bench (application must be running at localhost on port 8080).
It outputs the average time per iteration and the deviation between max and min. Example output from a 4 core 2.3GHz MacBook (application running inside docker limited to 4 cores and with 4 workers):
test bench_highhes ... bench: 71,112,344 ns/iter (+/- 7,798,699)
The application supports the following endpoints.
Signifies the application is healthy by returning a HTTP Status OK - 200 return code.
Prometheus formatted metrics. Currently exposes request count and duration per endpoint
Fetches and processes an image file. The only mandatory parameter is the image_address.
| Parameter | Description |
|---|---|
image_address |
The address for the Image. Should be a HTTP, HTTPS or HTTP valid URI. |
format |
desired image format. Possible values are Jpeg, Png, Heic and Webp. Defaults to Jpeg |
quality |
desired quality for the image. For Jpeg, it goes from 0 to 100 (defaults to 75) |
size[width] |
desired width for the image. Images won't get upscaled or have their aspect ratio changed by variations on parameters for width and height. |
size[height] |
desired height for the image. Images won't get upscaled or have their aspect ratio changed by variations on parameters for width and height. |
rotation |
optional rotation of the image. Possible values are R90, R180 and R270 |
Watermarks is an array parameter and therefore, must be indexed when informed (0 indexed).
| Parameter | Description |
|---|---|
watermarks[0][image_address] |
watermark file. File has to be smaller than original file. Should be a HTTP, HTTPS or HTTP valid URI. |
watermarks[0][alpha] |
opacity from the watermark over the original image. it is a floating point number from 0 to 1. |
watermarks[0][position][x][origin] |
identifier to position the watermark based on a point or centered (X axis). Possible values: Left (default), Right, Center. |
watermarks[0][position][y][origin] |
identifier to position the watermark based on a point or centered (Y axis). Possible values: Top (default), Bottom, Center. |
watermarks[0][position][x][pos] |
position of the watermark in the X axis. Value in pixels. |
watermarks[0][position][y][pos] |
position of the watermark in the Y axis. Value in pixels. |
watermarks[0][size] |
optional size of the watermark. It should be a value between 1 and 100 representing a percentage from the original image. |
(c) Copyright 2019-2020 OLX. Released under Apache 2 License