default.path.data included as a data path when path.data configured as yaml array

[tylerfontaine]

Elasticsearch version: 5.3.0

OS version: Tried in Centos 7 and OS X, though I suspect it doesn't matter.

Description of the problem including expected versus actual behavior:
When using multiple data paths, or even specifying just a single data path as an array in yaml, elasticearch includes whatever is passed to it as default.path.data, such as when installed via RPM or DEB packages, from the systemd unit OR the init.d script, since you set this when you execute the es binary:

-Edefault.path.data=${DATA_DIR}

Steps to reproduce:
Repro:

Set this in the elasticsearch.yml

path.data:
- "/some/datapath1"

run this:

bin/elasticsearch -Edefault.path.data="/some/datapath2"

And elasticsearch will start configured with both data paths.

Provide logs (if relevant):

[2017-04-07T13:30:49,987][INFO ][o.e.e.NodeEnvironment    ] [jCWIeyS] using [2] data paths, mounts [[/ (/dev/disk1)]], net usable_space [198.8gb], net total_space [464.7gb], spins? [unknown], types [hfs]

This also happens when you use this array format:

path.data: ["/some/datapath1"]

[jasontedor]

I know what is happening here and I see a fix, I will open a PR soon. This is an incredibly unfortunate consequence of the fact that when a setting is specified as an array, we break it down into new keys path.data.0, path.data.1, ..., path.data.n, yet the default setting is applied as path.data so none of the keys from the array setting override as is done with other default settings.

[jasontedor]

This is a horrifically bad bug, it appears to have been introduced in 5.3.0, and it means that shards can end up in /var/lib/elasticsearch even if a user did not intend for that to be the case because they configured path.data as an array but the packaging specifies default.path.data. If /var/lib/elasticsearch is sitting on the root partition because why not since the user wants the data elsewhere, the user can easily end up with:

• shards not where they wanted them
• the root partition can fill up

It gets worse. Let's say that we fix this bug in 5.3.1 and a user migrates from 5.3.0 not aware of the fiasco that is ensuing here. They start up Elasticsearch on 5.3.1, and now their shards are missing! If they're not careful, they are in data loss territory.

[nik9000]

😭

[nik9000]

Let's say that we fix this bug in 5.3.1 and a user migrates from 5.3.0 not aware of the fiasco that is ensuing here. They start up Elasticsearch on 5.3.1, and now their shards are missing! If they're not careful, they are in data loss territory.

If we just fix the config issue and someone upgrades one node at a time they shouldn't lose any data so long as they have replicas. The entire shard will have to recover from the primary, but that is a fairly normal thing. They'd have to manually clean up the shards in accidental data directory. Which isn't horrible, but it isn't great at all.

We could do more to try and copy the data or make tool or something.

Minimally anyone on 5.3 will want to be able to know whether they are effected by this issue before they upgrade so they know what they have to do as part of the upgrade.

[jasontedor]

If we just fix the config issue and someone upgrades one node at a time they shouldn't lose any data so long as they have replicas.

That's the problem, we do not have complete control over this.

They'd have to manually clean up the shards in accidental data directory. Which isn't horrible, but it isn't great at all.

Yes, if they are aware of the problem and do not panic. Or they do not do something like practice immutable infrastructure.

We could do more to try and copy the data or make tool or something.

Yes, but it's quite delicate.

Minimally anyone on 5.3 will want to be able to know whether they are effected by this issue before they upgrade so they know what they have to do as part of the upgrade.

Indeed, and therein lies the rub.

[nik9000]

Yeah, I just wanted to start talking about this. I certainly don't claim to have answers. I guess I'm fairly sure that the person doing the upgrade is going to want to know if they are effected by this. Lots of people won't and 5.3.0 will just be a normal version for them. For the effected people, even if we handle the whole thing automatically, they are going to need to know.

[jasontedor]

Yeah, I just wanted to start talking about this.

Indeed, we need more people thinking about this. I discussed some of the complexities with @dakrone earlier in the evening via another channel, and I've been thinking about it all night. Sorry that I haven't taken the time to write my thoughts down yet.

Another complexity that I thought of is that the fix for this has to live for the rest of the 5.x series lifecycle because someone can upgrade from 5.3.0 to any future release in the 5.x series.

And yet another complexity that I thought is a situation where someone has their default.path.data included as one of the values in path.data (assuming multiple path.data).

For the effected people, even if we handle the whole thing automatically, they are going to need to know.

I completely agree.

[tylerfontaine]

Do we have set limitations yet on rolling upgrades to 6.0? Would 5.3.0 -> 6.0 be allowed? If so, the fix has to carry over there as well.

…

On Fri, Apr 7, 2017, 8:58 PM Jason Tedor ***@***.***> wrote: Yeah, I just wanted to start talking about this. Indeed, I need more people to talk to about this. I discussed some of the complexities with @dakrone <https://github.com/dakrone> earlier in the evening via another channel, and I've been thinking about it all night. Sorry that I haven't taken the time to write my thoughts down yet. Another complexity that I thought of is that the fix for this has to live for the rest of the 5.x series lifecycle because someone can upgrade from 5.3.0 to any future release in the 5.x series. And yet another complexity that I thought is a situation where someone has their default.path.data included as one of the values in path.data (assuming multiple path.data). For the effected people, even if we handle the whole thing automatically, they are going to need to know. I completely agree. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#23981 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AB1ZYHwbhJ9d1nXtq3sn8vTdcT_x87O0ks5rtumtgaJpZM4M3P2Y> .