Comment by lxpz | Hacker News Reader

Comment by lxpz | Hacker News Reader

Parent Full thread

Losing a node is a regular occurrence, and a scenario for which Garage has been designed.

The assumption Garage makes, which is well-documented, is that of 3 replica nodes, only 1 will be in a crash-like situation at any time. With 1 crashed node, the cluster is still fully functional. With 2 crashed nodes, the cluster is unavailable until at least one additional node is recovered, but no data is lost.

In other words, Garage makes a very precise promise to its users, which is fully respected. Database corruption upon power loss enters in the definition of a "crash state", similarly to a node just being offline due to an internet connection loss. We recommend making metadata snapshots so that recovery of a crashed node is faster and simpler, but it's not required per se: Garage can always start over from an empty database and recover data from the remaining copies in the cluster.

To talk more about concrete scenarios: if you have 3 replicas in 3 different physical locations, the assumption of at-most one crashed node is pretty reasonable, it's quite unlikely that 2 of the 3 locations will be offline at the same time. Concerning data corruption on a power loss, the probability to lose power at 3 distant sites at the exact same time with the same data in the write buffers is extremely low, so I'd say in practice it's not a problem.

Of course, this all implies a Garage cluster running with 3-way replication, which everyone should do.

That is a much stronger guarantee than your documentation currently claims. One site falling over and being rebuilt without loss is great. One site losing power, corrupting the local state, then propagating that corruption to the rest of the cluster would not be fine. Different behaviours.

Fair enough, we will work on making the documentation clearer.

I think this is one where the behaviour is obvious to you but not to people first running across the project. In particular, whether power loss could do any of:

- you lose whatever writes to s3 haven't finished yet, if any

- the local node will need to repair itself a bit after rebooting

- the local node is now trashed and will have to copy all data back over

- all the nodes are now trashed and it's restore from backup time

I've been kicking the tyres for a bit and I think it's the happy case in the above, but lots of software out there completely falls apart on crashes so it's not generally a safe assumption. I think the behaviour is sqlite on zfs doesn't care about pulling the power cable out, lmdb is a bit further down the list.

So if you put a 3-way cluster in the same building and they lose power together, then what? Is your data toast?

If I make certain assumptions and you respect them, I will give you certain guarantees. If you don't respect them, I won't guarantee anything. I won't guarantee that your data will be toast either.

If you can't guarantee anything for all the nodes losing power at the same time, that's really bad.

If it's just the write buffer at risk, that's fine. But the chance of overlapping power loss across multiple sites isn't low enough to risk all the existing data.

I disagree that it's bad, it's a choice. You can't protect against everything. The team made calculations and decided that the cost to protect against this very low probability is not worth it. If all the nodes lose power you may have a bigger problem than that

Power outages across big areas are common enough.

It's downright stupid if you build a system that loses all existing data when all nodes go down uncleanly, not even simultaneously but just overlapping. What if you just happen to input a shutdown command the wrong way?

I really hope they meant to just say the write buffer gets lost.

That's why you need to go to other regions, not remain in the same area. Putting all your eggs in one basket (single area) _is_ stupid. Having a single shutdown command for the whole cluster _is_ stupid. Still accepting writes when the system is in a degraded state _is_ stupid. Don't make it sound worse than it actually is just to prove your point.

> Still accepting writes when the system is in a degraded state _is_ stupid.

Again, I'm not concerned for new writes, I'm concerned for all existing data from the previous months and years.

And getting in this situation only takes one out of a wide outage or a bad push that takes down the cluster. Even if that's stupid, it's a common enough stupid that you should never risk your data on the certainty you won't make that mistake.

You can't protect against everything, but you should definitely protect against unclean shutdown.

If it's a common enough occurrence to have _all_ your nodes down at the same time maybe you should reevaluate your deployment choices. The whole point of multi-nodes clustering is that _some_ of the nodes will always be up and running otherwise what you're doing is useless.

Also, garage gives you the possibility to automatically snapshot the metadata, advices on how to do the snapshotting at the filesystem level and to restore that.

All nodes going down doesn't have to be common to make that much data loss a terrible design. It just has to be reasonably possible. And it is. Thinking your nodes will never go down together is hubris. Admitting the risk is being realistic, not something that makes the system useless.

How do filesystem level snapshots work if nodes might get corrupted by power loss? Booting from a snapshot looks exactly the same to a node as booting from a power loss event. Are you implying that it does always recover from power loss and you're defending a flaw it doesn't even have?

No, the snapshotting and restore is manual

It sounds like that's a possibility, but why on earth would you take the time to setup a 3 node cluster of object storage for reliability and ignore one of the key tenants of what makes it reliable?

Reply on news.ycombinator.com