Current membership changing functionality of etcd seems to have a
problem which can cause deadlock.
How to produce:
1. construct N node cluster
2. add N new nodes with etcdctl member add, without starting the new members
What happens:
After finishing add N nodes, a total number of the cluster becomes 2 *
N and a quorum number of the cluster becomes N + 1. It means
membership change requires at least N + 1 nodes because Raft treats
membership information in its log like other ordinal log append
requests.
Assume the peer URLs of the added nodes are wrong because of miss
operation or bugs in wrapping program which launch etcd. In such a
case, both of adding and removing members are impossible because the
quorum isn't preserved. Of course ordinal requests cannot be
served. The cluster would seem to be deadlock.
Of course, the best practice of adding new nodes is adding one node
and let the node start one by one. However, the effect of this problem
is so serious. I think preventing the problem forcibly would be
valuable.
Solution:
This patch lets etcd forbid adding a new node if the operation changes
quorum and the number of changed quorum is larger than a number of
running nodes. If etcd is launched with a newly added option
-strict-reconfig-check, the checking logic is activated. If the option
isn't passed, default behavior of reconfig is kept.
Fixes https://github.com/coreos/etcd/issues/3477
It specifies request timeout error possibly caused by connection lost,
and print out better log for user to understand.
It handles two cases:
1. the leader cannot connect to majority of cluster.
2. the connection between follower and leader is down for a while,
and it losts proposals.
log format:
```
20:04:19 etcd3 | 2015-08-25 20:04:19.368126 E | etcdhttp: etcdserver:
request timed out, possibly due to connection lost
20:04:19 etcd3 | 2015-08-25 20:04:19.368227 E | etcdhttp: etcdserver:
request timed out, possibly due to connection lost
```
Before this PR, the timeout caused by leader election returns:
```
14:45:37 etcd2 | 2015-08-12 14:45:37.786349 E | etcdhttp: got unexpected
response error (etcdserver: request timed out)
```
After this PR:
```
15:52:54 etcd1 | 2015-08-12 15:52:54.389523 E | etcdhttp: etcdserver:
request timed out, possibly due to leader down
```
It uses heartbeat interval and election timeout to estimate the
commit timeout for internal requests.
This PR helps etcd survive under high roundtrip-time environment,
e.g., globally-deployed cluster.
Follow the simple rule in the atomic package:
"On both ARM and x86-32, it is the caller's responsibility to arrange
for 64-bit alignment of 64-bit words accessed atomically. The first word
in a global variable or in an allocated struct or slice can be relied
upon to be 64-bit aligned."
Tested on a system with /proc/cpuinfo reporting:
processor : 0
model name : ARMv7 Processor rev 1 (v7l)
Features : swp half thumb fastmult vfp edsp thumbee neon vfpv3
tls vfpv4 idiva idivt vfpd32 lpae evtstrm
CPU implementer : 0x41
CPU architecture: 7
CPU variant : 0x0
CPU part : 0xc0d
CPU revision : 1
After this PR, only cluster's interface Cluster is exposed, which makes
code much cleaner. And it avoids external packages to rely on cluster
struct in the future.
The PR extracts types.Cluster from etcdserver.Cluster. types.Cluster
is used for flag parsing and etcdserver config.
There is no need to expose etcdserver.Cluster public, which contains
lots of etcdserver internal details and methods. This is the first step
for it.
1. Persist the cluster version change through raft. When the member is restarted, it can recover
the previous known decided cluster version.
2. When there is a new leader, it is forced to do a version checking immediately. This helps to
update the first cluster version fast.
We store cluster related key in StoreAdminPrefix for some
historical reason. The previous API is called admin. But now,
the admin name is gone and `cluster` is a more clear and correct
name.
Cluster version is the min major.minor of all members in
the etcd cluster. Cluster version is set to the min version
that a etcd member is compatible with when first bootstrapp.
During a rolling upgrades, the cluster version will be updated
automatically.
For example:
```
Cluster [a:1, b:1 ,c:1] -> clusterVersion 1
update a -> 2, b -> 2
after a detection
Cluster [a:2, b:2 ,c:1] -> clusterVersion 1, since c is still 1
update c -> 2
after a detection
Cluster [a:2, b:2 ,c:2] -> clusterVersion 2
```
The API/raft component can utilize clusterVersion to determine if
it can accept a client request or a raft RPC.
We choose polling rather than pushing since we want to use the same
logic for cluster version detection and (TODO) cluster version checking.
Before a member actually joins a etcd cluster, it should check the version
of the cluster. Push does not work since the other members cannot push
version info to it before it actually joins. Moreover, we do not want our
raft RPC system (which is doing the heartbeat pushing) to coordinate cluster version.
The original process is stopping etcd only when pipeline message finds itself
has been removed. After this PR, stream dial has this functionality too.
It helps fast etcd stop, which doesn't need to wait for stream break to
fall back to pipeline, and wait for election timeout to send out message
to detect self removal.
Add remotes to rafthttp, who help newly joined members catch up the
progress of the cluster. It supports basic message sending to remote, and
has no stream connection for simplicity. remotes will not be used
after the latest peers have been added into rafthttp.
It is more reasonable to init the variable before passing it as an
argument.
It fixes a bug that etcdserver may panic on server stats when processing
a message from rafthttp streamReader before server stats is initialized
in server.Start().
It exposes the metrics of file descriptor limit and file descriptor used.
Moreover, it prints out warning when more than 80% of fd limit has been used.
```
2015/04/08 01:26:19 etcdserver: 80% of the file descriptor limit is open
[open = 969, limit = 1024]
```
Stop raftNode goroutine before stopping server goroutine, so
server.Stop does stop all underlying stuffs elegantly now. This fixes
the problem that previous-round lock on WAL may not be released when
etcd is restarted.
Do not aggressively compact raft log entries. After a snapshot,
etcd server can compact the raft log upto snapshot index. etcd server
compacts to an index smaller than snapshot to keep some entries in memory.
The leader can still read out the in memory entries to send to a slightly
slow follower. If all the entries are compacted, the leader will send the
whole snapshot or read entries from disk if possible.
Encoding store into json snapshot has quite high CPU cost. And it
will block for a while. This commit makes the encoding process non-
blocking by running it in another go-routine.
WAL should control the cut logic itself. We want to do falloc to
per allocate the space for a segmented wal file at the beginning
and cut it when it size reaches the limit.
* coreos/master:
rafthttp: fix import
raft: should not decrease match and next when handling out of order msgAppResp
Fix migration to allow snapshots to have the right IDs
add snapshotted integration test
fix test import loop
fix import loop, add set to types, and fix comments
etcdserver: autodetect v0.4 WALs and upgrade them to v0.5 automatically
wal: add a bench for write entry
rafthttp: add streaming server and client
dep: use vendored imports in codegangsta/cli
dep: bump golang.org/x/net/context
Conflicts:
etcdserver/server.go
etcdserver/server_test.go
migrate/snapshot.go
* coreos/master:
scripts: build-docker tag and use ENTRYPOINT
scripts: build-release add etcd-migrate
create .godir
raft: optimistically increase the next if the follower is already matched
raft: add handleHeartbeat handleHeartbeat commits to the commit index in the message. It never decreases the commit index of the raft state machine.
rafthttp: send takes raft message instead of bytes
*: add rafthttp pkg into test list
raft: include commitIndex in heartbeat
rafthttp: move server stats in raftHandler to etcdserver
*: etcdhttp.raftHandler -> rafthttp.RaftHandler
etcdserver: rename sender.go -> sendhub.go
*: etcdserver.sender -> rafthttp.Sender
Conflicts:
raft/log.go
raft/raft_paper_test.go
Compaction is now treated as an implementation detail of Storage
implementations; Node.Compact() and related functionality have been
removed. Ready.Snapshot is now used only for incoming snapshots.
A return value has been added to ApplyConfChange to allow applications
to track the node information that must be stored in the snapshot.
raftpb.Snapshot has been split into Snapshot and SnapshotMetadata, to
allow the full snapshot data to be read from disk only when needed.
raft.Storage has new methods Snapshot, ApplySnapshot, HardState, and
SetHardState. The Snapshot and HardState parameters have been removed
from RestartNode() and will now be loaded from Storage instead.
The only remaining difference between StartNode and RestartNode is that
the former bootstraps an initial list of Peers.
* coreos/master: (21 commits)
etcdserver: refactor ValidateClusterAndAssignIDs
integration: add integration test for remove member
integration: add test for member restart
version: bump to alpha.3
etcdserver: add buffer to the sender queue
*: gracefully stop etcdserver
Fix up migration tool, add snapshot migration
etcd4: migration from v0.4 -> v0.5
etcdserver: export Member.StoreKey
etcdserver: recover cluster when receiving newer snapshot
etcdserver: check and select committed entries to apply
etcdserver: recover from snapshot before applying requests
raft: not set applied when restored from snapshot
sender: support elegant stop
etcdserver: add StopNotify
etcdserver: fix TestDoProposalStopped test
etcdserver: minor cleanup
etcdserver: validate new node is not registered before in best effort
etcdserver: fix server.Stop()
*: print out configuration when necessary
...
Conflicts:
etcdserver/server.go
etcdserver/server_test.go
raft/log.go
Fixes all updates since bcwaldon sketched the original, with cleanup and
into an acutal working state. The commit log follows:
fix pb reference and remove unused file post rebase
unbreak the migrate folder
correctly detect node IDs
fix snapshotting
Fix previous broken snapshot
Add raft log entries to the translation; fix test for all timezones. (Still in progress, but passing)
Fix etcd:join and etcd:remove
print more data when dumping the log
Cleanup based on yichengq's comments
more comments
Fix the commited index based on the snapshot, if one exists
detect nodeIDs from snapshot
add initial tool documentation and match the semantics in the build script and main
formalize migration doc
rename function and clarify docs
fix nil pointer
fix the record conversion test
add migration to test suite and fix govet
Stop should be idempotent. It should simply send a stop signal to the server.
It is the server's responsibility to stop the go-routines and related components.
* coreos/master:
etcdserver: add sender tests
raft: Only call stableTo when we have ready entries or a snapshot.
etcdserver: add ID() function to the Server interface.
sender: use RoundTripper instead of Client in sender
* coreos/master: (27 commits)
pkg/wait: move wait to pkg/wait
etcdserver: do not add/remove/update local member to/from sender hub
etcdserver: not record attributes when add member
raft: add a test for proposeConfChange
raft: block Stop() on n.done, support idempotency
raft: add a test for node proposal
integration: add increase cluster size test
integration: remove unnecessary t.Testing argument
raft: stop the node synchronously
integration: fix test to propagate NewServer errors
etcdserver: move peer URLs check to config
etcdserver: ensure initial-advertise-peer-urls match initial-cluster
raft: add a test for node.Tick
raft: add comment string for TestNodeStart
etcdserver: use member instead of node at etcd level
raft: nodes return sorted ids
raft: update unstable when calling stableTo with 0
*: support updating advertise-peer-url Users might want to update the peerurl of the etcd member in several cases. For example, if the IP address of the physical machine etcd running on is changed, user need to update the adversite-pee-rurl accordingly. This commit makes etcd support updating the advertise-peer-url of its members.
transport: create a tls listener only if the tlsInfo is not empty and the scheme is HTTPS
etcdserver: use member pointer for all tests
...
Conflicts:
etcdserver/server.go
raft/log.go
raft/log_test.go
raft/node.go
Users might want to update the peerurl of the etcd member in several cases.
For example, if the IP address of the physical machine etcd running on is
changed, user need to update the adversite-pee-rurl accordingly.
This commit makes etcd support updating the advertise-peer-url of its members.
This change splits the raftLog.entries array into an in-memory
"unstable" list and a pluggable interface for retrieving entries that
have been persisted to disk. An in-memory implementation of this
interface is provided which behaves the same as the old version;
in a future commit etcdserver could replace the MemoryStorage with
one backed by the WAL.
There's no real need to expose a Discoverer interface/struct when the
only use of the interface (and indeed the module) is to invoke a single
function. This isn't Java, after all. So instead, simplify to Discovery
exposing just two functions: JoinCluster (i.e. what was formerly called
"discovery"), and GetCluster (hitherto "ProxyDiscovery")
Node set the applied to committed right after it sends out Ready to application. This is not
correct since the application has not actually applied the entries at that point. We add a
Advance interface to Node. Application needs to call Advance to tell raft Node its progress.
Also this change can avoid unnecessary copying when application is still applying entires but
there are more entries to be applied.
1. etcd fatals if there is critical error in the system and operator should
do something for it
2. etcd panics if there happens something unexpected, and it should be
reported to us to debug.
This creates a simple ID type (wrapped around uint64) to provide for
standard serialization/deserialization to a string (i.e. base 16
encoded). This replaces strutil so now that package is removed.
This is a simple solution to having the proxy keep up to date with the
state of the cluster. Basically, it uses the cluster configuration
provided at start up (i.e. with `-initial-cluster-state`) to determine
where to reach peer(s) in the cluster, and then it will periodically hit
the `/members` endpoint of those peer(s) (using the same mechanism that
`-cluster-state=existing` does to initialise) to update the set of valid
client URLs to proxy to.
This does not address discovery (#1376), and it would probably be better
to update the set of proxyURLs dynamically whenever we fetch the new
state of the cluster; but it needs a bit more thinking to have this done
in a clean way with the proxy interface.
Example in Procfile works again.
Continue using hex everywhere. Including here.
TODO: cleanup the printing of the structs which currently have decimal
to/from:
`{Type:MsgAppResp To:9973738105406047488 From:17050684879817348455 T...`
Removes the notion of name being anything more than advisory or
command-line grouping, and adds checks for bootstrapping the command
line. IDs are consistent if the URLs are consistent.
Xiang Li