- Fix: LookupN returns ring node destinations in ring order #211
- Fix: Use lower case sirupsen/logrus import #208
- Feature: Added GetNClients to router #212
- Allow requiresAppInPing to be configured #209
- Fix: Add option to enforce identical app name in pings #206
- Fix: Add WithError() to NoLogger #203
- Fix: remove error log during bootstrap complaining about not being bootstrapped #201
- Fix: ineffectual assignments #193
- Fix: Make lookups consistent on hash collisions #196
- Feature: Self-eviction #177
- Feature: Identity Carry Over #188 #189 #190 #191 #192 #194 #195
- Travis: Remove Go 1.5 support #187
Identity carry over provides a way to manually configure the identity of a member on the hashring. By changing the identity to be something else than its address, it becomes possible to guarantee ring equality before and after deploys in a dynamic environment (e.g. mesos).
By calling ringpop.SelfEvict
on shutdown, a member will declare itself as faulty
and gossips this to the other members.
This removes the time window where it would be marked as suspect
; in this way removing the time where other members
considered it to be part of the ring while it was not responding anymore.
Prior to ringpop v0.8.0 the address was used as the identity of a member. Starting with version v0.8.0, it's possible to
configure a separate identity. As a result, the behaviour of the Identity
and IdentityResolverFunc
has been changed.
The Identity
option now configures the identity of a member and will return an error when it matches an ip:port; services
that were using Identity
or IdentityResolverFunc
should now use the Address
and AddressResolverFunc
options.
You could use the following gofmt
snippets to easily refactor:
gofmt -r 'ringpop.Identity(a) -> ringpop.Address(a)' -w .
gofmt -r 'ringpop.IdentityResolverFunc(a) -> ringpop.AddressResolverFunc(a)' -w .
- Fix: Header leaking on the forwarding path caused calls to own endpoints to not be sharded when the same context was used while making the call #197
When applications use their incoming context blindly in an outgoing call it uses the same tchannel headers for the outgoing call as the incoming call had, causing incoming headers to be leaked in the outgoing call. Since most applications don't use headers there is nothing to worry about. However the forwarding path of ringpop uses a header to indicate a call has been forwarded. This header prevents indefinite forwarding loops to occur when applications fail to converge on ownership.
With this fix the header indicating a forwarded call will not be leaked to the application. This makes it possible for applications to use the incoming context on a call to a sharded endpoint on their own or another ringpop servers service an correctly route the request.
- Feature: Added label support to ringpop for nodes to annotate themselves with more information #167 #171 #172
- Maintainability: Refactored internal representation of member state for more flexible reincarnation and state change of a member #159 #161
- Fix: Make sure ringpop is listening for requests before bootstrapping #176 #146
- Fix: Added support for imports in
.thrift
files in generated code for thrift forwarding #162 - Fix: Mark generated code as being generated for suppression in diffs #169
- Fix: Be more specific in the functionality required from TChannel #166
- Tests: Run all tests on go 1.7 #179
- Tests: More stable unit tests, integration tests and race detector tests. Races are now mandatory tests on ci #164 #178 #181 #182
- Tests: All examples are tested on every pull request #157 #170
The ringpop interface changed two existing functions GetReachableMembers
and
CountReachableMembers
that now take a variadic argument of type
swim.MemberPredicate
instead of no arguments. This does not change the usage
of these functions, but does change the type of the function. This might cause
custom declared interfaces to not match ringpop anymore. The solution is to
change these functions in the interface used to match the current signature.
Previously the signature was:
GetReachableMembers() ([]string, error)
CountReachableMembers() (int, error)
The current signature is:
GetReachableMembers(predicates ...swim.MemberPredicate) ([]string, error)
CountReachableMembers(predicates ...swim.MemberPredicate) (int, error)
Due to a refactor in how event emitting is done the RegisterListener
method is
deprecated. Even though it still works and behaves as previously it will start
logging warnings. Since this code is not on the hot path only little log volume
is expected. Instead of this function it is now advised to use AddListener
.
This function also returns if the listener has been added or not.
- Fix: add 1-buffer for error channels, prevents leaking goroutines. #140
- Feature: support forwarding headers. #141
- Feature: testpop takes an additional -stats-udp or -stats-file flag that can be used to emit statsd messages to a file or over the network. #144
- Fix: race condition when loading hostlist in discover provider. #145
- Testing: add suspect, faulty and tombstope period command-line flags to testpop. #148
- Fix: use oneself as a source when creating a change for reincarnation. #150
- Maintainability: move from godep to Glide. #151
- Maintainability: emit lookup and lookupn stats and add a lookupn-event. #152.
- Fix: do not reincarnate twice unnecessarily. #153
- Feature: Automatic healing of fully partitioned rings that could be cause by temporary network failure. #135 #128 #129
- Fix: Fullsyncs are now executed in both directions to prevent persistent asymmetrical partitions. #134
- Fix: Overcome protocol incompatibilities in the dissemination of tombstones for the reaping of faulty nodes. #133 #132
- Fix: Leaking goroutines on an error response after a timeout. #140
- Maintainability: Refactor of internal use of the Discovery Provider. #130
- Testing: Pin mockery to a specific version for testing stability. #136
- Testing: Add Go race detector on CI. #137
It is advised to complete the deployment of this version within 24 hours after
the first node is upgraded. Since the reaping of faulty nodes is added to this
release a node that upgraded to this version will mark all the faulty members of
the ringpop cluster as tombstones (the new state introduced for the reaping of
faulty nodes) 24 hours after the deploy. If older versions of ringpop run in a
cluster that starts declaring these tombstones
the cluster will jump in endless
fullsyncs mode. Ringpop operates normally under these conditions but extra load
on both network and CPU are in effect during this time. The fullsyncs will
automatically resolve as soon as all members are upgraded to this version. Same
might happen during a partial rollback that operates for a longer period of time.
- Reverting reaping faulty nodes feature temporarily while investigate backwards compatibility issues.
- Feature: Faulty nodes are now automatically reaped from nodes' membership lists after (24 hours by default). #123
- New options for controlling suspect and reaping times. #123
- Add new
Ready
andDestroyed
events to ringpop. #125 - Add additional logging to bring ringpop-go on par with ringpop-node log messages. #116
- Fix bug where Ringpop automatically added itself to the bootstrap hosts for host-based bootstrapping, but not other bootstrapping methods. #120
- Fix race condition where membership and hashring could be inconsistent with each other. #112
- Remove
File
andHost
options from bootstrap options in favor ofDiscoverProvider
interface. #120 - Add Go 1.6 to testing on CI
Since 0.4.0 introduces a new node/member state, 0.4.0 is not backwards-compatible with previous versions.
Note rolling upgrades with older versions do work, but undefined behaviour will occur if two versions run in parallel for longer than the FaultyPeriod
(default 24 hours).
This release contains a breaking change to the options provided to the
ringpop.Bootstrap
call.
BootstrapOptions.File
and BootstrapOptions.Hosts
have been replaced with
BootstrapOptions.DiscoverProvider
. DiscoverProvider
is an interface which
requires a single method:
type DiscoverProvider interface {
Hosts() ([]string, error)
}
Ringpop comes with DiscoverProviders for the previous File
and Hosts
options out of the box.
To upgrade if you were previously using File
:
+ "github.com/uber/ringpop-go/discovery/jsonfile"
...
- bootstrapOpts.File = *hostfile
+ bootstrapOpts.DiscoverProvider = jsonfile.New(*hostfile)
For static Hosts
:
+ "github.com/uber/ringpop-go/discovery/statichosts"
...
- bootstrapOpts.Hosts = []string{"127.0.0.1:3000", "127.0.0.1:3001"}
+ bootstrapOpts.DiscoverProvider = statichosts.New("127.0.0.1:3000", "127.0.0.1:3001")
- Fix "keys have diverged" forwarding error for retries #69
- Fix possible race in disseminator #86
- Fix possible issue with leave state not being applied correctly #94
- Fix issues where unnecessary full syncs could occur #95, #97
- Improvements to join:
- Join retries now have exponential backoff #68
- Improved resilience to possible partitions at startup #65
- Reduce network chatter on join #85
- Hashring performance improvements (see discussion on #58) #58
- Revamped logging; new logger options #83
- New stats to aid partition detection #92, #104
- Improved test coverage across the board
- Update and test with latest TChannel (v1.0.3) #103
- Various fixes and improvements to test infrastructure
- Fix retry mechanism for forwarded requests. When multiple keys are forwarded and a retry is attempted, the retry would fail with a "key destinations have diverged" error.
- Fix goroutine leakage on forwarded requests that time out
- Add Thrift forwarding support #31
- Add
Ringpop.Ready()
#32 - Add
Ringpop.GetReachableMembers()
andRingpop.CountReachableMembers()
#29 - Add
Ringpop.Forward()
#26 - Lazy initialization and identity autodetection #40
- Improve stats and bring them in-line with those emitted by ringpop-node #46
- Automatically add self to bootstrap list #41
- Improve HashRing API #38
- Improve tests #35
- Improve constructor options pattern for Ringpop #33
- Disable TChannel retries and zipkin tracing #28
- Fix headers sent with forwarded requests #22
There are a significant number of breaking changes in this release:
- Ringpop constructor has been renamed from
Ringpop.NewRingpop
toringpop.New(app string, opts ...Option)
and now accepts optional functional arguments. See the package documentation for a list of options. ringpop.Bootstrap
now acceptsswim.BootstrapOptions
.ringpop.BootstrapOptions
has been removed.- Many public methods now return an error if they are called before the ring is bootstrapped. Signatures for the changed methods are:
Bootstrap(opts *swim.BootstrapOptions) ([]string, error)
Checksum() (uint32, error)
Lookup(key string) (string, error)
LookupN(key string, n int) ([]string, error)
Uptime() (time.Duration, error)
WhoAmI() (string, error)
- Removed public method
ringpop.Destroyed()
. Use the newringpop.Ready()
to determine if the ring is ready or not. - Events moved to
ringpop.events
package. This includes:EventListener
RingChangedEvent
RingChecksumEvent
LookupEvent
- The behaviour of creating a single-node cluster has changed. Calling
Bootstrap()
with no bootstrap file or hosts will now cause Ringpop to create a single-node cluster.
- Initial public release