CatgirlIntelligenceAgency/run

Run

When developing locally, this directory will contain run-time data required for the search engine. In a clean check-out, it only contains the tools required to bootstrap this directory structure.

Requirements

Docker - It is a bit of a pain to install, but if you follow this guide you're on the right track for ubuntu-like systems.

JDK 21 - The code uses Java 21 preview features. The civilized way of installing this is to use SDKMAN; graalce is a good distribution choice but it doesn't matter too much.

Set up

To go from a clean check out of the git repo to a running search engine, follow these steps. This assumes a test deployment. For a production like setup... (TODO: write a guide for this).

You're assumed to sit in the project root the whole time.

1. Run the one-time setup

It will create the basic runtime directory structure and download some models and data that doesn't come with the git repo because git deals poorly with large binary files.

$ run/setup.sh

2. Compile the project and build docker images

$ ./gradlew docker

3. Initialize the database

Before the system can be brought online, the database needs to be initialized. To do this, bring up the database in the background, and run the flyway migration tool.

$ docker-compose up -d mariadb
$ ./gradlew flywayMigrate

4. Bring the system online.

We'll run it in the foreground in the terminal this time because it's educational to see the logs. Add -d to run in the background.

$ docker-compose up

There are two docker-compose files available, docker-compose.yml and docker-compose-barebones.yml; the latter is a stripped down version that only runs the bare minimum required to run the system, for e.g. running a whitelabel version of the system. The former is the full system with all the frills of Marginalia Search, and is the one used by default.

To start the barebones version, run:

$ docker-compose -f docker-compose-barebones.yml up

5. You should now be able to access the system.

By default, the docker-compose file publishes the following ports:

Address	Description
http://localhost:8080/	User-facing GUI
http://localhost:8081/	Operator's GUI

Note that the operator's GUI does not perform any sort of authentication.
Preferably don't expose it publicly, but if you absolutely must, use a proxy or Basic Auth to add security.

6. Download Sample Data

A script is available for downloading sample data. The script will download the data from https://downloads.marginalia.nu/ and extract it to the correct location.

The system will pick the data up automatically.

$ run/download-samples.sh l

Four sets are available:

Name	Description
s	Small set, 1000 domains
m	Medium set, 2000 domains
l	Large set, 5000 domains
xl	Extra large set, 50,000 domains

Warning: The XL set is intended to provide a large amount of data for setting up a pre-production environment. It may be hard to run on a smaller machine and will on most machines take several hours to process.

The 'm' or 'l' sets are a good compromise between size and processing time and should work on most machines.

7. Process the data

Bring the system online if it isn't (see step 4), then go to the operator's GUI (see step 5).

• Go to Node 1 -> Storage -> Crawl Data
• Hit the toggle to set your crawl data to be active
• Go to Actions -> Process Crawl Data -> [Trigger Reprocessing]

This will take anywhere between a few minutes to a few hours depending on which data set you downloaded. You can monitor the progress from the Overview tab.

First the CONVERTER is expected to run; this will process the data into a format that can easily be inserted into the database and index.

Next the LOADER will run; this will insert the data into the database and index.

Next the link database will repartition itself, and finally the index will be reconstructed. You can view the process of these steps in the Jobs listing.

8. Run the system

Once all this is done, you can go to the user-facing GUI (see step 5) and try a search.

Important! Use the 'No Ranking' option when running locally, since you'll very likely not have enough links for the ranking algorithm to perform well.

Experiment Runner

The script experiment.sh is a launcher for the experiment runner, which is useful when evaluating new algorithms in processing crawl data.