Engineering Test Record
What We've Actually Tested
SpotWarp is new, so instead of quoting customers we don't have yet, here is the real test record — what we rented, what we broke on purpose, and what recovered. Every run below used real instances and real money on Vast.ai and RunPod.
2026-08-06 & 08-08
8 / 8 PASSED
Eviction & recovery, eight GPU tiers
Rented a real instance, started the real daemon, then destroyed the instance to force a genuine eviction. Recovery confirmed end to end on A100, H100, RTX 4090, A6000, L40S, V100, H200 and B200. Total real spend for the eight-tier day: under $1.
Real rentals, both clouds
ROUND TRIP
Cross-cloud bridge, verified in both directions
Not just Vast.ai → RunPod. We also verified the return leg and a full RunPod → Vast.ai → RunPod round trip on live rentals, so the bridge cloud is genuinely temporary rather than a one-way trip to a more expensive home.
2026-08-12
BYTE-FOR-BYTE
Backup integrity on a live pod
Wrote a checkpoint, a training log and a 5 MB weights file into /workspace on a real RunPod pod. All three arrived on the local machine inside a single sync cycle with matching contents, not just matching filenames.
Architecture change
20 MIN → < 60 S
Why recovery got fast
Trying replacement hosts one at a time meant one dead host could cost 10–20 minutes. SpotWarp now races four candidates concurrently and holds a 15-second grace window after the first responds, so the cheapest ready host wins — not merely the fastest to answer.
Found by testing
FIXED & SHIPPED
A bug our own testing caught
A live run landed on a Vast host that had failed the marketplace's own re-verification. It never connected, and quietly ran up a $4.90 bandwidth charge pulling a Docker image. SpotWarp now skips deverified hosts entirely. We mention it because this is the kind of thing you only find by actually renting machines.
Check it yourself
REPRODUCIBLE
You don't have to take our word
The client daemon is open source, the package is public on
PyPI, and every release binary ships with a
SHA256 checksum.
Read the failover code before you run it — that is the point of shipping it open.
What We've Actually Tested
SpotWarp is new, so instead of quoting customers we don't have yet, here is the real test record — what we rented, what we broke on purpose, and what recovered. Every run below used real instances and real money on Vast.ai and RunPod.
Eviction & recovery, eight GPU tiers
Rented a real instance, started the real daemon, then destroyed the instance to force a genuine eviction. Recovery confirmed end to end on A100, H100, RTX 4090, A6000, L40S, V100, H200 and B200. Total real spend for the eight-tier day: under $1.
Cross-cloud bridge, verified in both directions
Not just Vast.ai → RunPod. We also verified the return leg and a full RunPod → Vast.ai → RunPod round trip on live rentals, so the bridge cloud is genuinely temporary rather than a one-way trip to a more expensive home.
Backup integrity on a live pod
Wrote a checkpoint, a training log and a 5 MB weights file into
/workspaceon a real RunPod pod. All three arrived on the local machine inside a single sync cycle with matching contents, not just matching filenames.Why recovery got fast
Trying replacement hosts one at a time meant one dead host could cost 10–20 minutes. SpotWarp now races four candidates concurrently and holds a 15-second grace window after the first responds, so the cheapest ready host wins — not merely the fastest to answer.
A bug our own testing caught
A live run landed on a Vast host that had failed the marketplace's own re-verification. It never connected, and quietly ran up a $4.90 bandwidth charge pulling a Docker image. SpotWarp now skips deverified hosts entirely. We mention it because this is the kind of thing you only find by actually renting machines.
You don't have to take our word
The client daemon is open source, the package is public on PyPI, and every release binary ships with a SHA256 checksum. Read the failover code before you run it — that is the point of shipping it open.
And what we haven't done yet
We would rather you read this before subscribing than discover it afterwards. The 14-day money-back guarantee covers you either way.