a rough sketch for a communal task orchestrator built on AT
anirudh.fi
meshlet.md
edited
1(summarized by Claude)
2
3A decentralized, consent-based container orchestration system where commune members contribute compute resources and run each other's workloads. Built on AT infrastructure with mutual trust and fair-share principles.
4
5**Identity Model**
6- **Owner DIDs**: Real people (e.g., `did:plc:alice`)
7- **Meshlet DIDs**: Each meshlet gets its own DID (e.g., `did:plc:alice-mesh`)
8- **Two-way linkage**:
9 - Owner's repo: `commune.scheduler.meshlet { meshletDid: "did:plc:alice-mesh" }`
10 - Meshlet's repo: `commune.scheduler.meshletOwner { ownerDid: "did:plc:alice" }`
11
12**Cluster Formation**
13```bash
14meshlet --did=did:plc:alice-mesh \ # ← this would likely be implicit? maybe this could even just be a did:web:mesh.alice.com
15 --owner=did:plc:alice \
16 --cluster-members=did:plc:bob,did:plc:carol # owner DIDs
17```
18
19Cool thing about this is Alice's meshlet can be in multiple clusters -- insofar as the other meshlet nodes recognize her. There can be cluster overlap.
20
21Startup process:
221. For each member owner DID, resolve to their meshlet DID via their `commune.scheduler.meshlet` record
232. Verify mutual membership - check if their meshlet also lists your owner DID
243. Verify meshlet ownership - confirm meshlet→owner linkage
254. Form active cluster with only mutually-verified members
265. Members can participate in multiple overlapping clusters
27
28**Workload Scheduling (Decentralized)**
29
30No central scheduler - consensus via deterministic selection:
31
321. **Workload request**: Alice publishes `commune.scheduler.workload` to her owner repo
33 - Resource requirements (CPU, memory, storage)
34 - Container image
35 - Replica count
36 - Constraints
37
382. **Bid publication**: Meshlets see workload on firehose, evaluate capacity/fairness, publish `commune.scheduler.bid` to their meshlet repos
39 - Time window for bids (e.g., 5 seconds)
40 - Bids include priority score based on fair-share balance
41
423. **Deterministic selection**: All meshlets independently:
43 - Collect all bids from firehose
44 - Run identical selection algorithm (sort by priority + DID, take top N)
45 - Everyone reaches same conclusion about who runs it
46
474. **Execution**: Selected meshlets pull image and start containers
48
495. **Status updates**: Meshlets write `commune.scheduler.workloadStatus` to their repos
50
51**Communication Patterns**
52- **Firehose**: Workload discovery, bid coordination, status updates (all via atproto repos)
53- **XRPC**: Direct meshlet operations
54 - `farm.meshlet.getCapacity`
55 - `farm.meshlet.listWorkloads`
56 - `farm.meshlet.getWorkloadStatus`
57 - `farm.meshlet.deleteWorkload`
58 - etc. for pause/stop/shutdown meshlet etc.
59
60**Open Questions**
61- **Secrets management**: Sealed secrets per meshlet? Shared vault? Secret sharing?
62- **Fair-share accounting**: Local tracking vs distributed ledger?
63- **Failure handling**: What happens when meshlets go offline mid-workload?
64- **Governance**: Handling free-riders, minimum contributions, dispute resolution?
65- **Appview**: What would this look like?