Commit Graph
108 Commits
Author SHA1 Message Date
Matthias Radestock 27632b118f prevent excessive ring fragmentation due to repeated space requests
Merging a ring now indicates whether the ring has been updated. This
prevents excessive allocation retries, which in turn can result in
repeated space requests - i.e. if an allocation cannot proceed due to
lack of space - which in turn can result in excessive ring
fragmentation since many of the requests will result in the creation
of a new ring entries, due to the need to split ranges.

Fixes #2181.
2016-04-17 12:40:18 +01:00
Adam Harrison a09986baa2 Rename consense -> prime 2016-04-14 12:46:03 +01:00
Bryan Boreham 61b0d0ba9e Merge pull request #2149 from weaveworks/stop-test-router
stop ipam tests from leaking routers
2016-04-11 13:34:21 +01:00
Matthias Radestock 695babf7ba make Allocator.Stop() idempotent and non-blocking
So that tests don't have to be careful about ensuring that allocators
are stopped exactly once, or rely on the magic mesh.ChannelSize for
non-blocking behaviour.
2016-04-11 12:32:01 +01:00
Bryan Boreham ecf1887e5b Initialize Paxos quorum to 0 until this node needs to request consensus.
With a 0 quorum we will participate in voting, but will not actively
try to form a ring.
2016-04-11 11:42:24 +01:00
Bryan Boreham c12a52d20e Better log and status messages for Paxos observer 2016-04-07 12:50:52 +01:00
Bryan Boreham 0c59de40e7 Refactor Paxos status call to remove type assertion 2016-04-07 12:12:08 +01:00
Adam Harrison 436dc01f52 Merge pull request #2070 from weaveworks/issues/1721-quorum-callback
Use current target list to compute quorum
2016-03-29 11:11:30 +01:00
Bryan Boreham 2b617633ae Compute quorum based on the dynamic number of target peers 2016-03-24 12:06:59 +00:00
Tom Wilkie 7a7867ae16 Fix spelling mistakes. 2016-03-23 16:09:43 +00:00
Adam Harrison e9081c6607 Allow specification of initial IPAM ring via --ipam-seed 2016-03-21 09:49:54 +00:00
Matthias Radestock 9f10b21c9c use Space.NumFreeAddresses for calculating transferred ip count
...rather than Space.NumFreeAddressesInRange; we don't care about the
universe.

Alternatively we could make Space.AddRanges return the count. But
AddRanges doesn't know whether the ranges overlap with existing free
space - that is only guaranteed by the context - so would have to do
much the same before/after calculation.

Another alternative would be to add the range sizes in
AdminTakeoverRanges. But that is more code and less clear.
2016-03-19 14:41:36 +00:00
Matthias Radestock 62d823850c make rmpeer report on size of transferred space, and not fail
Closes #2058.
2016-03-19 14:40:26 +00:00
Matthias Radestock c48eefa82d refactor(ish): extract allocator logging commonality
and add warnf
2016-03-19 14:32:54 +00:00
Matthias Radestock 65eff683dd Merge branch '1.4' 2016-03-19 14:30:38 +00:00
Matthias Radestock f42e275884 broadcast ipam ring update after rmpeer
So that other peers notice the change more promptly.

Fixes #2068.
2016-03-19 13:57:20 +00:00
Bryan Boreham f11314ed3e Introduce 'Count' type to avoid abusing 'Offset' 2016-03-19 11:47:34 +00:00
Adam Harrison 2eeb754454 refactor: Introduce IPAM config object 2016-03-14 17:41:20 +00:00
Bryan Boreham a57335738d Don't do unnecessary clock skew check in IPAM 2016-03-10 11:41:47 +00:00
Adam Harrison 6584df274e Implement weave consense 2016-03-03 18:52:09 +00:00
Adam Harrison b84ef98de8 Support non-electing peers via --observer 2016-03-03 17:48:50 +00:00
Bryan Boreham 73f1a04d5f Deal in subnets, not ranges, when requesting address allocation 2016-03-02 17:21:51 -08:00
Bryan Boreham b76e2a87cf Store address+subnet prefix for 'owned' addresses 2016-03-02 17:21:24 -08:00
Bryan Boreham b6eec08247 Return all applicable results from Lookup
and stop erroring on Lookup not found; just return a blank result
2016-03-02 17:01:12 -08:00
Bryan Boreham 95cbc8940e Sync 'owned' data with existing containers at startup 2016-03-02 11:35:24 -08:00
Bryan Boreham f29716387f Persist the 'owned' address set 2016-03-02 11:35:24 -08:00
Bryan Boreham 7453c7a5dd Merge branch '1.4' 2016-02-29 08:22:22 -08:00
Adam Harrison bceabe32a2 Try all pending allocations
We can have outstanding allocations for different ranges owned by
different peers, so we should not short circuit after the first space
request.
2016-02-25 14:56:33 +00:00
Bryan Boreham 546f9592f9 Hard-fail on persistence errors, otherwise the network can get confused on a restart.
Create `Allocator.fatalf()`, for consistency
2016-02-12 18:12:26 +00:00
Bryan Boreham 06f3793f05 Remove persisted IPAM ring data on peerName change 2016-02-12 18:12:26 +00:00
Bryan Boreham f396c60402 Persist the Ring 2016-02-12 18:12:26 +00:00
Bryan Boreham ad036499eb Add db infra to IPAM 2016-02-12 18:12:26 +00:00
Bryan Boreham 07b79568dd Resurrect apparently-dead containers on ContainerStarted event 2016-02-04 13:49:43 +00:00
Matthias Radestock 6bebce0b5a tiny refactor 2016-02-02 14:03:01 +00:00
Bryan Boreham 71da517e88 Refactor to collect all code that touches alloc.owned 2016-02-02 13:57:12 +00:00
Bryan Boreham 952e9de265 Remove superfluous line and fix comment typo 2016-01-25 18:05:43 +00:00
Peter Bourgon e27b320ede Update import location for mesh 2016-01-15 15:57:01 +01:00
Bryan Boreham e468399616 Extend the timeout for dead containers to 30 seconds,
but remove them if they are destroyed, to reduce starvation
when users are creating and destroying containers quickly.
2016-01-12 14:38:44 +00:00
Bryan Boreham b2b293a8ee Added 'ContainerDestroyed()' callback from docker client 2016-01-12 14:38:44 +00:00
Bryan Boreham f84b933d7c Leave container allocations for a time after death
so they will be reclaimed if the container is restarted.
Still cancel pending operations immediately a container dies
2016-01-12 14:38:44 +00:00
Bryan Boreham 4672e9ec1f Generalise ticker so we can use it for more than paxos,
and call tryPendingOps() from it so we eventually re-try operations
2016-01-12 14:38:44 +00:00
Matthias Radestock 274265f001 typo 2015-12-21 17:03:03 +00:00
Matthias Radestock 3cd199e852 make GossipData immutable
The same piece of GossipData can be shared and subject to different
Merge operations. We either need to copy the data when we share it, or
create copies on modification, i.e. copy-on-write.

We choose the latter, since it is more efficient and there is just one
mutating operation.

Fixes #1697.
2015-11-17 20:13:46 +00:00
Matthias Radestock dc9f8762ca move generic communication substrate into its own 'mesh' package
This is mostly a matter of moving some files, changing package names,
changing some imports, and changing some type references.

We duplicate the odd constant & global var, and the intmac/macint
functions, since neither 'router' or 'mesh' nor, really,' common' are
a natural home for them.
2015-11-09 16:37:39 +00:00
Matthias Radestock 8bd77c0d47 refactor: get rid of router/utils.go
most functions end up in other files in the router package, but a few
land in common.
2015-11-07 15:45:29 +00:00
Bryan Boreham 67e9310cc1 Review feedback 2015-11-03 17:01:27 +00:00
Bryan Boreham 8742957f75 Tweak naming to be closer to code changes in nameserver 2015-11-03 14:15:49 +00:00
Bryan Boreham ace372cfc6 More extensive choice of transferee 2015-11-03 12:03:28 +00:00
Bryan Boreham b3cda39b0b Let IPAM pick any peer we know about to transfer ownership to, when shutting down 2015-11-02 18:34:58 +00:00
Tom Wilkie 33085e3798 Merge pull request #1499 from weaveworks/1463-ipam-seed-mismatch
detect when IPAM has been seeded by different peers
2015-10-09 14:46:55 +01:00