- 2021-08-20: Completed initial draft and distributed via a gist
- 2021-08-25: Migrated as an RFC and changed format
This document discusses the future of peer network management in Tendermint, with a particular focus on features, semantics, and a proposed roadmap. Specifically, we consider libp2p as a tool kit for implementing some fundamentals.
For the 0.35 release cycle the switching/routing layer of Tendermint was
replaced. This work was done "in place," and produced a version of Tendermint
that was backward-compatible and interoperable with previous versions of the
software. While there are new p2p/peer management constructs in the new
version (e.g. PeerManager and Router), the main effect of this change
was to simplify the ways that other components within Tendermint interacted with
the peer management layer, and to make it possible for higher-level components
(specifically the reactors), to be used and tested more independently.
This refactoring, which was a major undertaking, was entirely necessary to enable areas for future development and iteration on this aspect of Tendermint. There are also a number of potential user-facing features that depend heavily on the p2p layer: additional transport protocols, transport compression, improved resilience to network partitions. These improvements to modularity, stability, and reliability of the p2p system will also make ongoing maintenance and feature development easier in the rest of Tendermint.
The current (refactored) P2P stack is an improvement on the previous iteration (legacy), but as of 0.35, there remains room for improvement in the design and implementation of the P2P layer.
Some limitations of the current stack include:
- heavy reliance on buffering to avoid backups in the flow of components, which is fragile to maintain and can lead to unexpected memory usage patterns and forces the routing layer to make decisions about when messages should be discarded.
- the current p2p stack relies on convention (rather than the compiler) to enforce the API boundaries and conventions between reactors and the router, making it very easy to write "wrong" reactor code or introduce a bad dependency.
- the current stack is probably more complex and difficult to maintain because the legacy system must coexist with the new components in 0.35. When the legacy stack is removed there are some simple changes that will become possible and could reduce the complexity of the new system. (e.g. #6598.)
- the current stack encapsulates a lot of information about peers, and makes it difficult to expose that information to monitoring/observability tools. This general opacity also makes it difficult to interact with the peer system from other areas of the code base (e.g. tests, reactors).
- the legacy stack provided some control to operators to force the system to dial new peers or seed nodes or manipulate the topology of the system _in situ_. The current stack can't easily provide this, and while the new stack may have better behavior, it does leave operators hands tied.
Some of these issues will be resolved early in the 0.36 cycle, with the removal of the legacy components.
The 0.36 release also provides the opportunity to make changes to the protocol, as the release will not be compatible with previous releases.
These sections describe features that may make sense to include in a Phase 2 of a P2P project.
Currently, there's no provision for intranode communication using the P2P layer, which means when two reactors need to interact with each other they have to have dependencies on each other's interfaces, and initialization. Changing these interactions (e.g. transitions between blocksync and consensus) from procedure calls to message passing.
This is a relatively simple change and could be implemented with the following components:
- a constant to represent "local" delivery as the
Tofield onp2p.Envelope. - special path for routing local messages that doesn't require message serialization (protobuf marshalling/unmarshaling).
Adding these semantics, particularly if in conjunction with synchronous semantics provides a solution to dependency graph problems currently present in the Tendermint codebase, which will simplify development, make it possible to isolate components for testing.
Eventually, this will also make it possible to have a logical Tendermint node running in multiple processes or in a collection of containers, although the usecase of this may be debatable.
In the current system, all messages are sent with fire-and-forget semantics, and there's no coupling between a request sent via the p2p layer, and a response. These kinds of semantics would simplify the implementation of state and block sync reactors, and make intra-node message passing more powerful.
For some interactions, like gossiping transactions between the mempools of different nodes, fire-and-forget semantics make sense, but for other operations the missing link between requests/responses leads to either inefficiency when a node fails to respond or becomes unavailable, or code that is just difficult to follow.
To support this kind of work, the protocol would need to accommodate some kind
of request/response ID to allow identifying out-of-order responses over a
single connection. Additionally, expanded the programming model of the
p2p.Channel to accommodate some kind of _future_ or similar paradigm to
make it viable to write reactor code without needing for the reactor developer
to wrestle with lower level concurrency constructs.
Currently, all timeouts, buffering, and QoS features are handled at the router layer, and the reactors are implemented in ways that assume/require asynchronous operation. This both increases the required complexity at the routing layer, and means that misbehavior at the reactor level is difficult to detect or attribute. Additionally, the current system provides three main parameters to control quality of service:
- buffer sizes for channels and queues.
- priorities for channels
- queue implementation details for shedding load.
These end up being quite coarse controls, and changing the settings are difficult because as the queues and channels are able to buffer large numbers of messages it can be hard to see the impact of a given change, particularly in our extant test environment. In general, we should endeavor to:
- set real timeouts, via contexts, on most message send operations, so that senders rather than queues can be responsible for timeout logic. Additionally, this will make it possible to avoid sending messages during shutdown.
- reduce (to the greatest extent possible) the amount of buffering in channels and the queues, to more readily surface backpressure and reduce the potential for buildup of stale messages.
Currently the transport layer is message based, which makes sense from a mental model of how the protocol works, but makes it more difficult to implement transports and connection types, as it forces a higher level view of the connection and interaction which makes it harder to implement for novel transport types and makes it more likely that message-based caching and rate limiting will be implemented at the transport layer rather than at a more appropriate level.
The transport then, would be responsible for negotiating the connection and the
handshake and otherwise behave like a socket/file descriptor with Read and
Write methods.
While this was included in the initial design for the new P2P layer, it may be obviated entirely if the transport and peer layer is replaced with libp2p, which is primarily stream based.
In the current system, Tendermint assumes that all nodes in a network are largely equivalent, and nodes tend to be "chatty" making many requests of large numbers of peers and waiting for peers to (hopefully) respond. While this works and has allowed Tendermint to get to a certain point, this both produces a theoretical scaling bottle neck and makes it harder to test and verify components of the system.
In addition to peer's identity and connection information, peers should be able to advertise a number of services or capabilities, and node operators or developers should be able to specify peer capability requirements (e.g. target at least <x>-percent of peers with <y> capability.)
These capabilities may be useful in selecting peers to send messages to, it may make sense to extend Tendermint's message addressing capability to allow reactors to send messages to groups of peers based on role rather than only allowing addressing to one or all peers.
Having a good service discovery mechanism may pair well with the synchronous semantics (request/response) work, as it allows reactors to "make a request of a peer with <x> capability and wait for the response," rather force the reactors to need to track the capabilities or state of specific peers.
The current peer system is homegrown and is conceptually compatible with the needs of the project, and while there are limitations to the system, the p2p layer is not (currently as of 0.35) a major source of bugs or friction during development.
However, the current implementation makes a number of allowances for interoperability, and there are a collection of iterative improvements that should be considered in the next couple of releases. To maintain the current implementation, upcoming work would include:
- change the
Transportmechanism to facilitate easier implementations. - implement different
Transporthandlers to be able to manage peer connections using different protocols (e.g. QUIC, etc.) - entirely remove the constructs and implementations of the legacy peer implementation.
- establish and enforce clearer chains of responsibility for connection establishment (e.g. handshaking, setup,) which is currently shared between three components.
- report better metrics regarding the into the state of peers and network connectivity, which are opaque outside of the system. This is constrained at the moment as a side effect of the split responsibility for connection establishment.
- extend the PEX system to include service information so that nodes in the network weren't necessarily homogeneous.
While maintaining a bespoke peer management layer would seem to distract from development of core functionality, the truth is that (once the legacy code is removed,) the scope of the peer layer is relatively small from a maintenance perspective, and having control at this layer might actually afford the project with the ability to more rapidly iterate on some features.
LibP2P provides components that, approximately, account for the
PeerManager and Transport components of the current (new) P2P
stack. The Go APIs seem reasonable, and being able to externalize the
implementation details of peer and connection management seems like it could
provide a lot of benefits, particularly in supporting a more active ecosystem.
In general the API provides the kind of stream-based, multi-protocol supporting, and idiomatic baseline for implementing a peer layer. Additionally because it handles peer exchange and connection management at a lower level, by using libp2p it'd be possible to remove a good deal of code in favor of just using libp2p. Having said that, Tendermint's P2P layer covers a greater scope (e.g. message routing to different peers) and that layer is something that Tendermint might want to retain.
The are a number of unknowns that require more research including how much of a peer database the Tendermint engine itself needs to maintain, in order to support higher level operations (consensus, statesync), but it might be the case that our internal systems need to know much less about peers than otherwise specified. Similarly, the current system has a notion of peer scoring that cannot be communicated to libp2p, which may be fine as this is only used to support peer exchange (PEX,) which would become a property libp2p and not expressed in it's current higher-level form.
In general, the effort to switch to libp2p would involve:
- timing it during an appropriate protocol-breaking window, as it doesn't seem viable to support both libp2p and the current p2p protocol.
- providing some in-memory testing network to support the use case that the
current
p2p.MemoryNetworkprovides. - re-homing the
p2p.Routerimplementation on top of libp2p components to be able to maintain the current reactor implementations.
Open question include:
- how much local buffering should we be doing? It sort of seems like we should figure out what the expected behavior is for libp2p for QoS-type functionality, and if our requirements mean that we should be implementing this on top of things ourselves?
- if Tendermint was going to use libp2p, how would libp2p's stability guarantees (protocol, etc.) impact/constrain Tendermint's stability guarantees?
- what kind of introspection does libp2p provide, and to what extend would this change or constrain the kind of observability that Tendermint is able to provide?
- how do efforts to select "the best" (healthy, close, well-behaving, etc.) peers work out if Tendermint is not maintaining a local peer database?
- would adding additional higher level semantics (internal message passing, request/response pairs, service discovery, etc.) facilitate removing some of the direct linkages between constructs/components in the system and reduce the need for Tendermint nodes to maintain state about its peers?