Pinned post
Anicet
Verified account boat@AniC_dev
captain of boat, the best full VM sandbox @boatdotdev
prev: HPC for black holes hunting @esa, gamedev
prev: HPC for black holes hunting @esa, gamedev
1.7K Following 12K Followers 18K Posts
Latest posts 5
our resume cold starts / filesystem snapshot restore should now feel way faster
we recently had a period of extremely fast growth, notably thanks to @theo, that caused some stability and speed issues, I'm deeply sorry for it
thankfully our VMs remained up >99.95% of the time
but we experienced severe slowdowns on filesystem snapshot restore speed due to congestion on the storage and poor optimization
happy to announce the situation is now stable, fixes have been shipped and the speed is back to normal, but we still want to cut it massively soon!
now also VM state is cached for some time to allow instant fast stop -> resume cycles, typically great for chat apps/personal assistants
graphs below show you uncached full restore times, to compare with other providers please consider their uncached times as well (not their "I resume state in 20ms" claims which are cached times)
our stop/resume and snapshot system is built on object storage and FUSE. it allows us to store your files for cheap, for long, at scale, and with extremely good persistence & restorability guarantees (we don't lose anything even for months)
as you edit files in your VM the changes are saved gradually in the background, even if you don't explicitely ask for a snapshot
and when you resume we get you a fresh VM, mount the whole disk via FUSE so files appear to be virtually there instantly, while we download them from storage
also we never want you to lose work:
if the VM fails to snapshot and you stop it, we stop billing you, but keep the VM up, retry snapshotting, email you, and if we persist we manually intervene
---
this is all just the tip of the iceberg, there are still ways to cut restore time by multiples, and remove a bunch of bugs that remain, @luaroncrew is on it
we recently had a period of extremely fast growth, notably thanks to @theo, that caused some stability and speed issues, I'm deeply sorry for it
thankfully our VMs remained up >99.95% of the time
but we experienced severe slowdowns on filesystem snapshot restore speed due to congestion on the storage and poor optimization
happy to announce the situation is now stable, fixes have been shipped and the speed is back to normal, but we still want to cut it massively soon!
now also VM state is cached for some time to allow instant fast stop -> resume cycles, typically great for chat apps/personal assistants
graphs below show you uncached full restore times, to compare with other providers please consider their uncached times as well (not their "I resume state in 20ms" claims which are cached times)
our stop/resume and snapshot system is built on object storage and FUSE. it allows us to store your files for cheap, for long, at scale, and with extremely good persistence & restorability guarantees (we don't lose anything even for months)
as you edit files in your VM the changes are saved gradually in the background, even if you don't explicitely ask for a snapshot
and when you resume we get you a fresh VM, mount the whole disk via FUSE so files appear to be virtually there instantly, while we download them from storage
also we never want you to lose work:
if the VM fails to snapshot and you stop it, we stop billing you, but keep the VM up, retry snapshotting, email you, and if we persist we manually intervene
---
this is all just the tip of the iceberg, there are still ways to cut restore time by multiples, and remove a bunch of bugs that remain, @luaroncrew is on it
Twitter is just a bunch of SF people copying that one French guy's best ideas
Twitter is just a bunch of SF people copying each other words IRL and posting them.View quoted post on X
