FA-99296 / Distributed training aggregation / Open access
Retransmitted replica updates contribute more than once · case 01
Retransmitted replica updates contribute more than once.
ROOT CAUSE
Collective retries can deliver the same worker and step more than once.
THE FAILURE
Collective retries can deliver the same worker and step more than once.
Unsuccessful approach: Deduplicating by worker alone suppresses legitimate contributions from another step.
Case contract
Rows [step,worker,value], target step; last row for each worker at that step wins, then values sum.
Why this case matters
A toy distributed-training contract makes replica and shard contribution rules explicit without claiming framework compatibility or performance.
1 / The failure
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(rows, step):
return sum(v for s,w,v in rows if s==step)
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
r=[[N,'a',N],[N,'a',3*N],[N,'b',2*N]]
check('retry replaces',solve(r,N),5*N)
check('empty',solve([],N),0)
check('other step',solve([[N+1,'a',9*N]],N),0)
check('same worker different steps',solve([[N-1,'a',100],[N,'a',N]],N),N)
check('singleton',solve([[N,'a',4]] ,N),4)
check('last zero retry',solve([[N,'a',N],[N,'a',0]],N),0)
check('stale retry after current',solve([[N,'a',N],[N-1,'a',50*N]],N),N)
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)
| Boundary fixture | Actual | Expected | Outcome |
|---|---|---|---|
| retry replaces | 6 | 5 | Failed |
| empty | 0 | 0 | Passed |
| other step | 0 | 0 | Passed |
| same worker different steps | 1 | 1 | Passed |
| singleton | 4 | 4 | Passed |
| last zero retry | 1 | 0 | Failed |
| stale retry after current | 1 | 1 | Passed |
SHA-256 / 96421dcb384a4b6a2006a2087ca3041949d39e58e57d4abf69a8dba857d88ea5
2 / The unsuccessful fix
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(rows, step):
d={}
for s,w,v in rows: d[w]=(s,v)
return sum(v for s,v in d.values() if s==step)
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
r=[[N,'a',N],[N,'a',3*N],[N,'b',2*N]]
check('retry replaces',solve(r,N),5*N)
check('empty',solve([],N),0)
check('other step',solve([[N+1,'a',9*N]],N),0)
check('same worker different steps',solve([[N-1,'a',100],[N,'a',N]],N),N)
check('singleton',solve([[N,'a',4]] ,N),4)
check('last zero retry',solve([[N,'a',N],[N,'a',0]],N),0)
check('stale retry after current',solve([[N,'a',N],[N-1,'a',50*N]],N),N)
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)
| Boundary fixture | Actual | Expected | Outcome |
|---|---|---|---|
| retry replaces | 5 | 5 | Passed |
| empty | 0 | 0 | Passed |
| other step | 0 | 0 | Passed |
| same worker different steps | 1 | 1 | Passed |
| singleton | 4 | 4 | Passed |
| last zero retry | 0 | 0 | Passed |
| stale retry after current | 0 | 1 | Failed |
SHA-256 / 3d4c8373d7b0346a8e44a37001b04f5d60bb73cbb7563bb16e7d16184396cae9
HELD IN THE MEMBER ARCHIVE
The verified repair and its recorded checks are member-only.
This mechanism has 7 recorded checks per implementation. The open-access tier publishes the failure and the unsuccessful fix; the repaired source that passes every check, and the observations that prove it, are available to members.
Every case sharing this mechanism uses the same contract and the same repair, so this one record is held back for all of them.
Member access is invitation-based. Sign in with your invited account to inspect the repair.
Sign in to the archive ↗Verification & scope
Inputs are already materialized Python values from a fixed toy collective; no transport, failure recovery, optimizer state, or real model behavior is represented. This reproducer isolates one failure mechanism. Results cover the supplied fixtures. Variants within a family share a test contract and should remain grouped when constructing evaluation splits. Related mechanisms with a shared evaluation_group must also remain together; these controlled models are not independent production incidents.
Observations recorded using Python 3.12.14 at 2026-09-30T16:48:22.633137+00:00.
Case digest / 70ae321b5225d806940a88be4a583791b3083751906fd7575db60886bb0e921d