Long story: Since #1768 there has been a recurring repinning test failure with
Raft consensus.
Per the test, if a pin is allocated to a peer that has been shutdown,
submitting the pin again should re-allocate it to a peer that is still
running.
Investigation on why this test fails and why it fails only in Raft lead to
realizing that this and other similar tests, were passing by chance. The
needed re-allocations were made not by the new submission of the pin, but by
the automatic-repinning feature. The actual resubmitted pin was carrying the
same allocations (one of them being the peer that was down), but it was
silently failing because the RedirectToLeader() code path was using
cc.ctx and hitting the peer that had been shutdown, which caused it to error.
Fixing the context propagation, meant that we would re-overwrite the pin with
the old allocations, thus the actual behaviour did not pass the test.
So, on one side, this fix an number of tests that had not disabled automatic
repinning and was probably getting in the way of things. On the other side,
this removes a condition that prevents re-allocation of pins if they exists
and options have not changed.
I don't fully understand why this was there though, since the Allocate() code
does return the old allocations anyways when they are enough, so it should not
re-allocate randomly. I suspect this was preventing some misbehaviour in the
Allocate() code from the time before it was improved with multiple allocators
etc.
We are propagating the wrong context (mostly from the Cluster top-level
methods). This makes that request cancellations (and cancellations of the
associated contexts) are not propagated to many methods, and can result in
deadlocks when an operation that is holding a lock is not aborted.
This affects for example the operation tracker. Getting all operations from
the tracker relies on someone reading from the out channel, or on the context
being cancelled. When a request is aborted in the middle of the response, and
the context is not cancelled, everything that wants to list operations would
become deadlocked, including operations that need write locks like
TrackNewOperation.
This fixes it.
This commit introduces the new go-libp2p-gorpc streaming capabilities for
Cluster. The main aim is to work towards heavily reducing memory usage when
working with very large pinsets.
As a side-effect, it takes the chance to revampt all types for all public
methods so that pointers to static what should be static objects are not used
anymore. This should heavily reduce heap allocations and GC activity.
The main change is that state.List now returns a channel from which to read
the pins, rather than pins being all loaded into a huge slice.
Things reading pins have been all updated to iterate on the channel rather
than on the slice. The full pinset is no longer fully loaded onto memory for
things that run regularly like StateSync().
Additionally, the /allocations endpoint of the rest API no longer returns an
array of pins, but rather streams json-encoded pin objects directly. This
change has extended to the restapi client (which puts pins into a channel as
they arrive) and to ipfs-cluster-ctl.
There are still pending improvements like StatusAll() calls which should also
stream responses, and specially BlockPut calls which should stream blocks
directly into IPFS on a single call.
These are coming up in future commits.
* Libp2p protectors no longer needed, use PSK directly
* Generate cluster 32-byte secret here (helper gone from pnet)
* Switch to go-log/v2 in all places
* DHT bootstrapping not needed. Adjust DHT options for tests.
* Do not rely on dissappeared CidToDsKey and DsKeyToCid functions fro dshelp.
* Disable QUIC (does not support private networks)
* Fix tests: autodiscovery started working properly
TrustedPeers are specified in the configuration. Additional peers
can be added at runtime with Trust/Distrust functions.
Unfortunately we cannot use consensus.PeerAdd as a way to trust a peer as
cluster.PeerAdd+Join can be called by any peer and this calls
consensus.PeerAdd.
The result is consensus.PeerAdd doing a lot in Raft while consensus.Trust does
nothing, while in CRDTs consensus.Trust does something but consensus.PeerAdd
does nothing. But this is more or less consistent.
I had thought of this for a very long time but there were no compelling
reasons to do it. Specifying RPC endpoint permissions becomes however
significantly nicer if each Component is a different RPC Service. This also
fixes some naming issues like having to prefix methods with the component name
to separate them from methods named in the same way in some other component
(Pin and IPFSPin).
This adds a new "crdt" consensus component using go-ds-crdt.
This implies several refactors to fully make cluster consensus-component
independent:
* Delete mapstate and fully adopt dsstate (after people have migrated).
* Return errors from state methods rather than ignoring them.
* Add a new "datastore" modules so that we can configure datastores in the
main configuration like other components.
* Let the consensus components fully define the "state.State". Thus, they do
not receive the state, they receive the storage where we put the state (a
go-datastore).
* Allow to customize how the monitor component obtains Peers() (the current
peerset), including avoiding using the current peerset. At the moment the
crdt consensus uses the monitoring component to define the current peerset.
Therefore the monitor component cannot rely on the consensus component to
produce a peerset.
* Re-factor/re-implementation of "ipfs-cluster-service state"
operations. Includes the dissapearance of the "migrate" one.
The CRDT consensus component defines creates a crdt-datastore (with ipfs-lite)
and uses it to intitialize a dssate. Thus the crdt-store is elegantly
wrapped. Any modifications to the state get automatically replicated to other
peers. We store all the CRDT DAG blocks in the local datastore.
The consensus components only expose a ReadOnly state, as any modifications to
the shared state should happen through them.
DHT and PubSub facilities must now be created outside of Cluster and passed in
so they can be re-used by different components.
This takes advantange of the latest features in go-cid, peer.ID and
go-multiaddr and makes the Go types serializable by default.
This means we no longer need to copy between Pin <-> PinSerial, or ID <->
IDSerial etc. We can now efficiently binary-encode these types using short
field keys and without parsing/stringifying (in many cases it just a cast).
We still get the same json output as before (with minor modifications for
Cids).
This should greatly improve Cluster performance and memory usage when dealing
with large collections of items.
License: MIT
Signed-off-by: Hector Sanjuan <hector@protocol.ai>
This commit adds support for OpenCensus tracing
and metrics collection. This required support for
context.Context propogation throughout the cluster
codebase, and in particular, the ipfscluster component
interfaces.
The tracing propogates across RPC and HTTP boundaries.
The current default tracing backend is Jaeger.
The metrics currently exports the metrics exposed by
the opencensus http plugin as well as the pprof metrics
to a prometheus endpoint for scraping.
The current default metrics backend is Prometheus.
Metrics are currently exposed by default due to low
overhead, can be turned off if desired, whereas tracing
is off by default as it has a much higher performance
overhead, though the extent of the performance hit can be
adjusted with smaller sampling rates.
License: MIT
Signed-off-by: Adrian Lanzafame <adrianlanzafame92@gmail.com>
Unfortunately, there are still some data races in yamux
https://github.com/libp2p/go-libp2p/issues/396 so we can't
enable this by default.
License: MIT
Signed-off-by: Hector Sanjuan <code@hector.link>
This commit promotes the Consensus component (and Raft) to become a fully
independent thing like other components, passed to NewCluster during
initialization. Cluster (main component) no longer creates the consensus
layer internally. This has triggered a number of breaking changes
that I will explain below.
Motivation: Future work will require the possibility of running Cluster
with a consensus layer that is not Raft. The "consensus" layer is in charge
of maintaining two things:
* The current cluster peerset, as required by the implementation
* The current cluster pinset (shared state)
While the pinset maintenance has always been in the consensus layer, the
peerset maintenance was handled by the main component (starting by the "peers"
key in the configuration) AND the Raft component (internally)
and this generated lots of confusion: if the user edited the peers in the
configuration they would be greeted with an error.
The bootstrap process (adding a peer to an existing cluster) and configuration
key also complicated many things, since the main component did it, but only
when the consensus was initialized and in single peer mode.
In all this we also mixed the peerstore (list of peer addresses in the libp2p
host) with the peerset, when they need not to be linked.
By initializing the consensus layer before calling NewCluster, all the
difficulties in maintaining the current implementation in the same way
have come to light. Thus, the following changes have been introduced:
* Remove "peers" and "bootstrap" keys from the configuration: we no longer
edit or save the configuration files. This was a very bad practice, requiring
write permissions by the process to the file containing the private key and
additionally made things like Puppet deployments of cluster difficult as
configuration would mutate from its initial version. Needless to say all the
maintenance associated to making sure peers and bootstrap had correct values
when peers are bootstrapped or removed. A loud and detailed error message has
been added when staring cluster with an old config, along with instructions on
how to move forward.
* Introduce a PeerstoreFile ("peerstore") which stores peer addresses: in
ipfs, the peerstore is not persisted because it can be re-built from the
network bootstrappers and the DHT. Cluster should probably also allow
discoverability of peers addresses (when not bootstrapping, as in that case
we have it), but in the meantime, we will read and persist the peerstore
addresses for cluster peers in this file, different from the configuration.
Note that dns multiaddresses are now fully supported and no IPs are saved
when we have DNS multiaddresses for a peer.
* The former "peer_manager" code is now a pstoremgr module, providing utilities
to parse, add, list and generally maintain the libp2p host peerstore, including
operations on the PeerstoreFile. This "pstoremgr" can now also be extended to
perform address autodiscovery and other things indepedently from Cluster.
* Create and initialize Raft outside of the main Cluster component: since we
can now launch Raft independently from Cluster, we have more degrees of
freedom. A new "staging" option when creating the object allows a raft peer to
be launched in Staging mode, waiting to be added to a running consensus, and
thus, not electing itself as leader or doing anything like we were doing
before. This additionally allows us to track when the peer has become a
Voter, which only happens when it's caught up with the state, something that
was wonky previously.
* The raft configuration now includes an InitPeerset key, which allows to
provide a peerset for new peers and which is ignored when staging==true. The
whole Raft initialization code is way cleaner and stronger now.
* Cluster peer bootsrapping is now an ipfs-cluster-service feature. The
--bootstrap flag works as before (additionally allowing comma-separated-list
of entries). What bootstrap does, is to initialize Raft with staging == true,
and then call Join in the main cluster component. Only when the Raft peer
transitions to Voter, consensus becomes ready, and cluster becomes Ready.
This is cleaner, works better and is less complex than before (supporting
both flags and config values). We also backup and clean the state whenever
we are boostrapping, automatically
* ipfs-cluster-service no longer runs the daemon. Starting cluster needs
now "ipfs-cluster-service daemon". The daemon specific flags (bootstrap,
alloc) are now flags for the daemon subcommand. Here we mimic ipfs ("ipfs"
does not start the daemon but print help) and pave the path for merging both
service and ctl in the future.
While this brings some breaking changes, it significantly reduces the
complexity of the configuration, the code and most importantly, the
documentation. It should be easier now to explain the user what is the
right way to launch a cluster peer, and more difficult to make mistakes.
As a side effect, the PR also:
* Fixes#381 - peers with dynamic addresses
* Fixes#371 - peers should be Raft configuration option
* Fixes#378 - waitForUpdates may return before state fully synced
* Fixes#235 - config option shadowing (no cfg saves, no need to shadow)
License: MIT
Signed-off-by: Hector Sanjuan <code@hector.link>
The --version flag is default from our cli library so I left that. The
version subcommand prints only the version number + the short commit
so it's a bit more easy to parse.
I have additionally reduced the amount of output on start up by converting
some messages to debug. I wish there was a level between INFO and DEBUG
though.
License: MIT
Signed-off-by: Hector Sanjuan <code@hector.link>
This is what it was likely causing PeerRemove tests to fail randomly
but very often. We cancelled the Cluster context before shutting down
the Consensus component. This killed networking and aborted
the peer remove operations when the leader is removing itself.
As a result, it would error with "leadership lost", which would
trigger a retry which would set the final error to "context cancelled"
because the shutdown of the consensus component proceeds during the
retry, cancelling the consensus context.
This is not only affecting tests, it might affected operations when
running cluster.
License: MIT
Signed-off-by: Hector Sanjuan <hector@protocol.ai>
I think this will prevents some random tests failures
when we realize that we are not anymore in the peerset
and trigger a shutdown but Raft has not finished fully
committing the operation, which then triggers an error,
and a retry. But the contexts are cancelled in the retry
so it won't find a leader and will error finally error
with that message.
License: MIT
Signed-off-by: Hector Sanjuan <hector@protocol.ai>
This change removes the duplicities of the PeerManager component:
* No more commiting PeerAdd and PeerRm log entries
* The Raft peer set is the source of truth
* Basic broadcasting is used to communicate peer multiaddresses
in the cluster
* A peer can only be added in a healthy cluster
* A peer can be removed from any cluster which can still commit
* This also adds support for multiple multiaddresses per peer
License: MIT
Signed-off-by: Hector Sanjuan <hector@protocol.ai>
The main differences is that the new version of Raft is more strict
about starting raft peers which already contain configurations.
For a start, cluster will fail to start if the configured cluster
peers are different from the Raft peers. The user will have to
manually cleanup Raft (TODO: an ipfs-cluster-service command for it).
Additionally, this commit adds extra options to the consensus/raft
configuration section, adds tests and improves existing ones and
improves certain code sections.
License: MIT
Signed-off-by: Hector Sanjuan <hector@protocol.ai>
The following commit reimplements ipfs-cluster configuration under
the following premises:
* Each component is initialized with a configuration object
defined by its module
* Each component decides how the JSON representation of its
configuration looks like
* Each component parses and validates its own configuration
* Each component exposes its own defaults
* Component configurations are make the sections of a
central JSON configuration file (which replaces the current
JSON format)
* Component configurations implement a common interface
(config.ComponentConfig) with a set of common operations
* The central configuration file is managed by a
config.ConfigManager which:
* Registers ComponentConfigs
* Assigns the correspondent sections from the JSON file to each
component and delegates the parsing
* Delegates the JSON generation for each section
* Can be notified when the configuration is updated and must be
saved to disk
The new service.json would then look as follows:
```json
{
"cluster": {
"id": "QmTVW8NoRxC5wBhV7WtAYtRn7itipEESfozWN5KmXUQnk2",
"private_key": "<...>",
"secret": "00224102ae6aaf94f2606abf69a0e278251ecc1d64815b617ff19d6d2841f786",
"peers": [],
"bootstrap": [],
"leave_on_shutdown": false,
"listen_multiaddress": "/ip4/0.0.0.0/tcp/9096",
"state_sync_interval": "1m0s",
"ipfs_sync_interval": "2m10s",
"replication_factor": -1,
"monitor_ping_interval": "15s"
},
"consensus": {
"raft": {
"heartbeat_timeout": "1s",
"election_timeout": "1s",
"commit_timeout": "50ms",
"max_append_entries": 64,
"trailing_logs": 10240,
"snapshot_interval": "2m0s",
"snapshot_threshold": 8192,
"leader_lease_timeout": "500ms"
}
},
"api": {
"restapi": {
"listen_multiaddress": "/ip4/127.0.0.1/tcp/9094",
"read_timeout": "30s",
"read_header_timeout": "5s",
"write_timeout": "1m0s",
"idle_timeout": "2m0s"
}
},
"ipfs_connector": {
"ipfshttp": {
"proxy_listen_multiaddress": "/ip4/127.0.0.1/tcp/9095",
"node_multiaddress": "/ip4/127.0.0.1/tcp/5001",
"connect_swarms_delay": "7s",
"proxy_read_timeout": "10m0s",
"proxy_read_header_timeout": "5s",
"proxy_write_timeout": "10m0s",
"proxy_idle_timeout": "1m0s"
}
},
"monitor": {
"monbasic": {
"check_interval": "15s"
}
},
"informer": {
"disk": {
"metric_ttl": "30s",
"metric_type": "freespace"
},
"numpin": {
"metric_ttl": "10s"
}
}
}
```
This new format aims to be easily extensible per component. As such,
it already surfaces quite a few new options which were hardcoded
before.
Additionally, since Go API have changed, some redundant methods have been
removed and small refactoring has happened to take advantage of the new
way.
License: MIT
Signed-off-by: Hector Sanjuan <hector@protocol.ai>
