Repository navigation
RFC 0002: elastic execution — worker pools, provider governor, tiered models - #260
Open
pd-admin-jc wants to merge 3 commits into
Open
pd-admin-jc wants to merge 3 commits into
pd-admin-jc wants to merge 3 commits into
Conversation
Contributor
There was a problem hiding this comment.
🟡 Changes recommended
The RFC contains unresolved scaling, quota-accounting, concurrency, and model-routing design flaws.
11 open findings
What changed in this PR
Defines RFC 0002 for elastic execution in JarvisCore 2.2.
Changes:
- Proposes elastic worker pools and distributed backlog scaling.
- Designs provider-aware rate governance and priority lanes.
- Introduces per-turn model tiers with acceptance criteria.
| File | Description |
|---|---|
rfcs/0002-elastic-execution.md |
Documents the elastic execution architecture and delivery plan. |
🧠 Review effort: Balanced
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
|
|
||
| - **Durable leased work.** `RedisContextStore.claim_step` leases a step atomically to | ||
| one execution identity. Leases expire and are recovered, and committing requires | ||
| holding the lease. Exactly-once execution and resume after a crash depend on this. |
| Each process runs one scaler coroutine per pool, once a second: | ||
|
|
||
| ``` | ||
| desired = clamp(ceil(backlog / per_worker), min, max) |
|
|
||
| ``` | ||
| desired = clamp(ceil(backlog / per_worker), min, max) | ||
| desired = min(desired, governor.headroom_workers(pool.model_tier)) |
Comment on lines
+112
to
+114
| - **Across processes:** members register in `pool_members:{capability}` with a | ||
| heartbeat TTL. The global `max` is enforced there, so two API replicas don't | ||
| each start 32. |
Comment on lines
+138
to
+140
| 1. **Provider metadata, when credentials allow.** For Azure, the deployment's | ||
| `sku.capacity` from Resource Manager is the TPM in thousands. RPM follows the | ||
| documented ratio of 6 RPM per 1,000 TPM for standard deployments. |
|
|
||
| ### 5.2 Admission | ||
|
|
||
| Every call reserves capacity before it is dispatched: |
| cost = estimate(prompt_tokens) + max_tokens # what the provider charges at admission | ||
| reserve(key, requests=1, tokens=cost, lane) -> admitted | wait(seconds) | ||
| ... call ... | ||
| settle(key, reservation, actual_total_tokens) # returns the unused estimate |
Comment on lines
+189
to
+192
| Lanes are `interactive` (chat turns), `decision` (final work products, verdicts) | ||
| and `bulk` (research loops). Bulk is admitted only while headroom stays above a | ||
| reserve, by default 15 % of TPM. A user's chat question during a 196-claim run is | ||
| therefore not queued behind it. |
Comment on lines
+211
to
+214
| - **Loop turns** use the fast tier: choosing tools, issuing searches, reading, | ||
| extracting quotes. | ||
| - **The final turn** uses the strong tier: the turn that emits `DONE/RESULT` for a | ||
| work-product contract. |
| |---|---|---| | ||
| | Governor (§5) | automatic | automatic: every profile calls the same `self.llm` client | | ||
| | Pools (§4) | `mesh.add(..., pool=Pool(...))` | the same, for agents that keep no per-instance state | | ||
| | Tiered turns (§6) | automatic: the kernel knows which turn is final | per call: `self.llm.generate(..., tier="fast" \| "strong")`; omit it for the default | |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
RFC 0002: elastic execution. One agent scales across pool members while its work queues, a provider governor paces every model call per deployment, and tiered models let cheap turns run on a cheaper model.
Design goals:
Implementation PRs:
feat/provider-governor: the provider governor (PR 1).feat/elastic-pools: elastic pools, stacked on the governor (PR 2).