Meilisearch deployment on Fargate

[wperron]

This PR introduces the necessary AWS resources to deploy Meilisearch on Fargate. I personally like Fargate because while the straight compute-hour cost is higher that EC2, it removes a lot of complexity from the management side of things; logging and metrics monitoring is baked-in with Container Insights, the ECS service takes care of restarting failed instances automatically without needing an Autoscaling Group, using a Docker image means we don't have to deal with managing server configurations or ssh into a server. I'm not religiously attached to Fargate, we can discuss the pros/cons and I always move to a more traditional EC2 deployment if need be.

Here's a rough price estimate:

• Fargate: $21.62 with a single instance at 0.5vcpu, 2Gb memory
• EFS: < $0.30 (less than 1Gb total)
• ALB: $16.74/mo flat rate + LCU-hours depending on new http connections per hour

Note that the Load Balancer isn't stricly necessary, we could also use a simple dns-based round-robin and get the job done just fine; ALB simply makes it easier to integrate with Security Groups and AWS Certificate Manager.

Checklist:

• perform some tests to make sure the EFS volume mount works as expected
• perform some tests with more than one instance (load & resiliency)
• perform tests with meili version upgrades
• configure Meilisearch master key to control access
• tweak LB to only allow external traffic to the /search endpoint
• adjust resource allocation based on the amount of data we estimate we'll need
• get some metrics on the current usage of the search endpoint to estimate the LB costs
• Implement a backup strategy

Fixes #69

[wperron]

The EFS volume works great; dropping and recreating a task attaches to the volume. Also interesting point; dropping the ecs service actually waits for Meili to empty its internal queue before terminating cleanly, I'll try this out by dropping a task directly instead of the overarching service just to make sure.

[wperron]

Running with 3 instances, I get some weird artifacts. I think it's related to how Meili updates its internal state, but I would create a "modules" index and then curl /indexes/modules , and depending on the instance hit, I would get a mix of 404s and 200s for at least 5 minutes. Search results hits would also vary depending on the instance hit by the request.

I also did the same test locally just to be sure it wasn't because of the EFS mount, and I get the same behavior locally.

For now I think running with a single instance is a better option, we can revisit the multi-instance setup when the cluster mode lands in meili.

[wperron]

After consideration, I don't think Fargate is the right choice right now due to both the low amount of data we have and how much harder Fargate has been so far; EC2 just seems like a much simpler option.