feat(scheduling): replica spread + resource-aware auto placement (#56)

services.spread distributes replicas across every online, schedulable
agent (replica i → agents[i mod n], deterministic). auto placement now
orders by real headroom — memory utilisation, cpu usage, then container
count — instead of count alone.

Lifecycle (start/stop/restart/delete) and GET /runtime are now driven
by container_instances inventory, so remote replicas are visible and
tear down wherever they run — covers spread, pins, and stale pin
leftovers. Spread and explicit pins are mutually exclusive.

Generated with [Devin](https://devin.ai)

Co-authored-by: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
Tomáš DvořákandDevin authored and GitHub committed 2026-10-09 17:03:26 +02:00
1 parent bc93c83cf9
commit dd5ce80ce2
21 files changed
+469 -100

No files matched your search

+10 -3
View File
@@ -289,11 +289,18 @@ and wired refs, one click or one CLI call.
(drain evicts remote containers via the agent queue + unpins services
so they redeploy locally), `nodes cordon|uncordon|drain` CLI +
MCP (drain confirm-gated) + node-detail page controls; `auto` and
explicit pins skip cordoned nodes. Remote nodes run registry-pulled
explicit pins skip cordoned nodes. Spread placement shipped:
`services.spread` distributes replicas deterministically across every
online schedulable agent (replica i → agents[i mod n]), mutually
exclusive with a pin; `auto` is now resource-aware — least memory
utilisation, then cpu, then container count, with uninstrumented
agents sorting last. Lifecycle and `GET /runtime` are
inventory-driven, so remote replicas show up and tear down wherever
they run. Remote nodes run registry-pulled
images only — git builds, Traefik domains, and remote sleep are
documented ceilings. Remaining: `scheduling_rules` spread/affinity
documented ceilings. Remaining: `scheduling_rules` affinity
policies, deployment status reported back from agent
results, per-node resource-aware `auto` (currently count-based).
results, spread weighting by real capacity.
- [🟡] **Build-on-node** — private registry pulls shipped: the dispatch
payload carries the project owner's `registries` credentials for the
image host; the agent `docker login`s for the pull and logs out after,