SN-DBS distribution probe

SN-DBS is a distributed build farm (dbsutil, dbsrun, dbsbuild on PATH via a version-independent shim in C:\Program Files (x86)\SCE\Common\SN-DBS\bin). Use this to get evidence about what it will and will not distribute, instead of reasoning from the manual.

1. Confirm the network and inventory

dbsutil -connected     # "SN-DBS network is available" + active broker
dbsutil -l             # every agent: Agent: "name" fqdn (ip) vX.Y [enabled]

dbsutil -l is an authoritative fleet inventory — better than AD computer accounts, because an enabled agent is a live machine with a running service. Parse with: ^Agent:\s+"(?<name>[^"]+)"\s+\S+\s+\((?<ip>[^)]+)\)\s+\S+\s+\[(?<state>[^\]]+)\]

2. Flood, never trickle

A small batch proves nothing: SN-DBS runs jobs locally when local CPU is free, and the report says so with "where":"local", local_scheduling_reasons: {"allow_local_cpu": ""}. Post ~300 jobs so work spills onto agents.

Script file = one command per line (batch style; && chains within a line). Then:

dbsrun dbsbuild -p <ProjectName> -gt --report <out.json> -s <script.bat>

-gt allows generic tools (no tool template). --report is the whole point: it writes JSON with completed_jobs[], each carrying where (local|remote), host_name, exit_code, stdout, stderr, failovers.

3. Read the report, not the console

import json, collections
d = json.loads(open('out.json').read())
jobs = d['completed_jobs']
print(collections.Counter(j['where'] for j in jobs))
print(collections.Counter(j.get('host_name') or 'LOCAL' for j in jobs))
print(collections.Counter(j['exit_code'] for j in jobs))

Have the probe program print its own hostname so stdout independently proves execution on the remote box, not just shipping.

4. Known-good controls

• hostname.exe — native, on every box: baseline that distribution works at all.
• C:\Windows\System32\curl.exe -s -o NUL -w "http=%{http_code} connect=%{time_connect}" <url> — proves DNS + TCP egress from inside the sandbox to your target.

5. Rules that constrain any design

• Native PE only. Normal .NET fails remotely with The application to execute does not exist: '<app>.dll'. NativeAOT works (PublishAot=true, single file, no side-by-side DLLs).
• No machine targeting. No affinity attribute exists; the scheduler picks. Exclusions are GUI-only (Network View → Members → right-click), not scriptable — only DedicatedBuildServer appears in dbsutil -list-settings.
• Batch semantics. Jobs must run to completion, honour a kill timeout, take input from argv/env/files, and return output via stdout/stderr/files. One failure can abort the batch — exit 0 whenever you produced output.
• Idle-only agents that withdraw when their owner gets busy: fine for throughput work, wrong for pinned-concurrency work.

6. Design shape that fits

Emit one line of JSON per job on stdout (counts, sums, and a capped reservoir of raw samples), then merge the pooled samples driver-side. Never average per-job percentiles.