[frectrium]

What

Adds docs/design/exercise-plan.md: what we build in Week 3, where it runs and what we measure.

• Erasure coding: RS(6,3) on 9 nodes, 3 per zone, so a whole zone can be lost.
• New data: 3 synchronous replicas, one per zone, then closed 64 MiB volumes are coded in the background.
• Environments: 9-node Compose cluster for development. For measurement, 9 m7g.medium nodes, one m7g.large control instance and a c7g.large load generator in ap-south-1.
• Experiments: E1 latency vs block size against S3, E2 scaling, E3 replication vs erasure coding, E4 node and zone failure, E5 resource usage for the cost model. 3 repeats each.
• Plumbing: values for [Infra] Terraform: 3-AZ Satchel cluster in ap-south-1 with one-command teardown #20, [Infra] Observability: Prometheus metrics and Grafana dashboards #21, [Experiment] Load generator and S3 baseline harness #22, [Docs] Test strategy #31, [Infra] Integration test harness on the local cluster, run in CI #32, [Infra] AWS account and Terraform groundwork #33 and [Infra] Publish the Satchel Docker image for deployments #34, and the list of build issues for Create Week 3 build and experiment issues from the exercise plan #19.

Closes #14

Why

Week 3 can't start without fixed parameters. Each decision mirrors Magic Pocket and is recorded with its sources in the Design Decision Log (#17), so this doc gives the DDL-NNN ID instead of repeating it.

How to check it

• Section 1: one row per question in [Design] Finalize the exercise plan, including the erasure coding scheme and parameters #14, each with its DDL entry.
• Section 4: the experiment matrix and the instance-hour estimate (about 17 hours of an 11-instance cluster). Tell me if it's too big for the budget.
• Section 7: the on-prem estimate ([Cost] Build the on-premises cost estimate #25) uses Magic Pocket's 3.0x overhead from the workload model ([Cost] Build the case workload model from Dropbox's public disclosures #15), not Satchel's layout.
• Section 8: the numbers the test strategy ([Docs] Test strategy #31) is waiting for.

Decisions to confirm in review

• RS(6,3) on 9 nodes, 3 synchronous replicas before coding
• SQLite instead of MySQL for the index
• 64 MiB volumes, closed after 10 minutes
• The Master in the verifier process
• Instance types and the size of the experiment matrix

Merge order

Merge after the Design Decision Log PR (#55). Once this merges, I create the Week 3 build issues (#19).

Checklist

• cargo fmt, cargo clippy and cargo test pass locally (code PRs)
• Tests added or updated for the change (code PRs)
• New sources are added to docs/sources/README.md (research PRs)
• Design decisions are recorded in the Design Decision Log
• Docs and README are updated if behaviour or setup changed