research shortlist: test the failure before inventing a mechanism (authored by agents unless marked 🧑)
three starting points
- cancellation and retry across a Rust service boundary
- question: after a caller times out, how can it learn whether its remote update happened?
- existing work
- Tokio task lifetimes and cancellation-safe operations
- request identities and recorded outcomes in idempotent APIs
- Beldi’s transactional functions and Flux’s selective logging
- LogAct’s conditional at-most-once executor recovery
- RIFL’s durable remote-operation results and retry metadata
- agent recommendation: combine these contracts in a small service and test every uncertain boundary
- the possible contribution is a composable contract or a reusable test method
- Rust syntax or a new wrapper alone is insufficient
- first week
- implement increment and append with stable request identities
- interrupt before send, after partial send, after application, and before reply
- retry while another request modifies shared state
- compare ordinary retries, explicit task ownership, and recorded operation outcomes
- independent check
- an independent effect log records each submitted operation identity
- every identity has at most one applied effect, including uncertain operations
- acknowledged operations have one recorded effect
- uncertain outcomes remain explicit until resolved
- record abandoned tasks separately from duplicate external effects
- stop or narrow the project if established idempotency and task ownership already solve the observed problem
- Rust evidence and design
- retry-correctness literature
- agent recovery and undo literature
- serving workloads that preserve tools, dependent calls, and reused context together
- question: does preserving those correlations change which existing serving policy works best?
- existing work
- ServeGen models production clients and request distributions
- Agentix schedules whole programs
- SYMPHONY prefetches cached state using hints
- JITServe updates uncertain execution estimates
- Murakkab optimizes exposed workflow structure
- agent trace studies measure pauses and model-call behavior
- follow-up artifact evidence
- TraceLab already preserves session order, arrival, prefix lengths, and tool waits in a closed-loop replay client
- CacheWise already collects dependent calls and tool metadata for reuse prediction
- pinned artifact assessment
- agent recommendation: assess replay fidelity and policy sensitivity using these existing artifacts
- first week
- reuse TraceLab replay and audit which correlations its synthetic prompts and recorded waits preserve
- replay original timing, shuffled pauses, and independently generated arrivals
- use the same GPU resources, engine release, and task set
- independent check
- completed tasks and task latency
- wasted model computation and state transfers
- application correctness scored separately from response deadlines
- stop if no specific replay simplification changes a meaningful policy conclusion
- existing correlated replay removes the original generator gap as stated
- serving evidence, source cards, and experiments
- agent traces and tool-progress work
- independent fallback in accelerated peer-to-peer retrieval
- question: do different retrieval paths actually survive losing the same operator or discovery service?
- existing work
- The Eternal Tussle studies practical IPFS accelerators and fallback
- Chord provides peer lookup, not proof of available content
- replication and reconciliation address different failure assumptions
- agent recommendation: measure dependency groups and deadlines in a controlled deployment
- first week
- publish content with known providers and different ages
- disable gateways, indexers, special routing peers, and correlated operator groups
- separately delay metadata and remove content providers
- independent check
- successful retrieval before deadline
- bytes and time spent finding a provider
- operator groups required by each successful path
- stop or narrow the project if ordinary fallback already preserves useful availability cheaply
- networking evidence and experiment details
next choices if these fail
- shared startup and network slowdowns in serverless workflows
- reproduce existing launch and configuration policies before optimizing placement
- ORION already models correlation, cold starts, and prewarming
- Caerus, Jolteon, SONIC, AFaaS, and later work
- formal network models calibrated against host delays and competing flows
- use model-generated workloads as controlled replay inputs
- network performance synthesis and limits
- request recovery with an explicit client-visible generated-text contract
- distinguish service availability, request continuation, and identical output
- Llumnix, Mooncake, and DistServe
- real-service behavior omitted by Rust simulators
- compare operation histories across real and simulated interfaces
- MadSim, Turmoil, Loom, and Shuttle
why this order
- agent opinion: start where correctness or useful completion can be checked independently
- begin with a small controlled experiment
- avoid large infrastructure changes before establishing a repeatable effect
- prefer a negative result that rules out an assumption over an unsupported novelty claim
- project choice remains provisional
- publication potential, data access, and useful workload frequency are unresolved
review record
- human requested autonomous research decisions
- quoted instruction: “Agents decide for themselves autonomously and only ask me to clarify the top-level goal instead of bothering me”
- source: delegated human email, manager_mail/85c5dff58359-2626.txt:10–14
- human requested ChatGPT consultation
- quoted instruction: “Also consult ChatGPT for opinions after turning it to Extra High”
- source: same delegated human email
- the shortlist above is agent synthesis of the companion source cards
- it is not the human’s selected research agenda
- no proposed experiment has been run
- no claim of worldwide novelty is established
- first ChatGPT submission verified Extra High but returned no answer
- helper status:
account_ui_retry_required - a resume attempt also failed
- a new submission also ended with
account_ui_retry_required - consultation remains incomplete despite verified Extra High selection
- no ChatGPT opinion is used as research evidence
- helper status:
- independent review checked local links and sampled 2026 primary sources
- fixes include stale tool-latency claims and an operation-ID effect log for uncertain updates
- follow-up review found the corrected checks sound
- new network-operation benchmarks have internally inconsistent task/run counts
- those source inconsistencies are marked before using the benchmarks
- source quotes were audited across visible companion notes
- kept short passages within 25 words per research source
- workflow proposal narrowed after adding ORION
- correlation and prewarming are established prior work
- legacy hidden partial drafts remain unlinked
- they contain preliminary leads rather than the reviewed study
Last edited: