Cache old partition states incase of ZooKeeper failure. #65
Conversation
partitions.go
Outdated
| replication: replication, | ||
| local: make(map[int]bool), | ||
| remote: make(map[int][]string), | ||
| old: make(map[int][]string, 1024), |
There was a problem hiding this comment.
old isn't really a descriptive name. maybe disappeared?
partitions.go
Outdated
| } | ||
| } | ||
|
|
||
| // Keep track of old peers in case zookeeper goes away. |
There was a problem hiding this comment.
This comment isn't particularly helpful. I think it'd be better to add documentation to FindPeers about the new functionality.
partitions.go
Outdated
| for _, node := range unDedupedPartition { | ||
| if !found[node] { | ||
| found[node] = true | ||
| p.old[partitionId] = append([]string{node}, p.old[partitionId]...) |
There was a problem hiding this comment.
It'd probably be slightly more efficient to append rather than prepend, and then truncate as part of the dedupe (working backwards). Even better, though, would be to just have a separate dedupe method.
There was a problem hiding this comment.
Also, this always adds the nodes to this list, when we just want to add them if they got removed above.
partitions.go
Outdated
|
|
||
| oldPeers := make([]string, 1024) | ||
| copy(oldPeers, p.old[partition]) | ||
| // Append old peers to peer list, in case of Zookeeper issues. |
There was a problem hiding this comment.
Again, I don't think this comment tells the reader much (if they don't already know what's up).
partitions.go
Outdated
| oldPeers := make([]string, 1024) | ||
| copy(oldPeers, p.old[partition]) | ||
| // Append old peers to peer list, in case of Zookeeper issues. | ||
| peers = append(peers, oldPeers...) |
There was a problem hiding this comment.
This means the dead peers will get shuffled into the live peers by the caller of this code. I think it'd be better if FindPeers return shuffled live peers + dead peers at the end.
partitions.go
Outdated
| unDedupedPartition := append(newPartition, p.old[partitionId]...) | ||
| found := map[string]bool{} | ||
|
|
||
| // Shitty dedupe, iterate though the remote peers |
partitions.go
Outdated
| } | ||
| } | ||
|
|
||
| func (p *partitions) deDupe(nodes []string) []string { |
There was a problem hiding this comment.
also there's no reason for it to be on p: it can just be func dedupe(s []string) []string
partitions.go
Outdated
|
|
||
| func (p *partitions) deDupe(nodes []string) []string { | ||
| found := map[string]bool{} | ||
| dedupedNodes := make([]string, len(nodes)) |
There was a problem hiding this comment.
this should be make([]string, 0, len(nodes)) or you're going to have extra empty values at the end
partitions.go
Outdated
| } | ||
| } | ||
|
|
||
| for partitionId, partition := range remote { |
There was a problem hiding this comment.
this still isn't checking that the remote peer actually disappeared (ie, that it used to be in p.remote and now is not)
partitions.go
Outdated
| } | ||
|
|
||
| for partitionId, partition := range remote { | ||
| newPartition := make([]string, len(partition)) |
There was a problem hiding this comment.
I'd call this peers I think
partitions.go
Outdated
| newPartition := make([]string, len(partition)) | ||
| copy(newPartition, partition) | ||
|
|
||
| unDedupedPartition := append(newPartition, p.disappeared[partitionId]...) |
There was a problem hiding this comment.
you don't need the temp variable here.
partitions.go
Outdated
| return path.Join(p.zkPath, fmt.Sprintf("%05d@%s", partition, p.peers.address)) | ||
| } | ||
|
|
||
| // getPeers returns the list of peers who have the given partition available. |
There was a problem hiding this comment.
You need to update the docstring. However, I still think it'd be cleaner to just return one list, and consider the peers "ordered" by priority. That gives us the freedom to actually add in priority later, if we want to do some light downweighting of peers from which we've seen errors.
|
Speaking of the build failing, this would be great with some tests! |
scottjab-stripe
force-pushed
the
scottjab/cache-old-peers-for-a-bit
branch
from
c8be9fe to
e18c85e
Compare
sharding/partitions.go
Outdated
| return p | ||
| } | ||
|
|
||
| // Dedupelicates elements in a slice of strings. |
There was a problem hiding this comment.
This should be of the form dedupe deduplicates...
sharding/partitions.go
Outdated
| } | ||
| } | ||
|
|
||
| // FindPeers returns the list of peers who have the given partition available. |
There was a problem hiding this comment.
This docstring is still wrong, and I still think the signature would be better if it just returned a list of peers vaguely prioritized. That way we could trivially add some sort of greylisting for peers that exhibit errors a lot.
serve.go
Outdated
| } | ||
|
|
||
| func shuffle(vs []string) []string { | ||
| func shuffle(vs []string, disappeared []string) []string { |
There was a problem hiding this comment.
This will return an arbitrary number of deleted nodes in addition to the live ones, which will change the behavior of proxy subtly. I think we want to try just peers, where dead peers are only considered if there aren't enough live peers to fulfill the replication factor.
2 participants
Sequins will start 501ing if zookeeper has a blip. This PR attempts to cache partition state between blips.
r? @colinmarc
cc? @charleseff