How to handle networking/peer APIs
This article applies to the code in git master only
Introduction
The bitcoinj networking APIs have a few options targeted at different use-cases - you can spin up individual Peers and manage them yourself or bring up a PeerGroup
to let it manage them, you can use one-off sockets or socket managers, and you can use blocking sockets or NIO/non-blocking sockets. This page attempts to explain the tradeoffs and use-cases for each choice, as well as provide some basic examples to doing more advanced networking.
The bitcoinj networking API is built up in a series of layers. On the bottom are simple wrapper classes that provide an API to open new connections using blocking sockets or java NIO (asynchronous select()-based sockets). On top of those sit various parsers that parse the network traffic into messages (ie into the Bitcoin messages). On top of those are Peer
objects, which handle message handling (exchanging initial version handshake, downloading blocks, etc) for each individual remote peer and provide a simple event listener interface. Finally a PeerGroup
can be layered on top to keep track of peers, ensuring there are always enough connections to the network and keeping track of network sync progress.
The simple case
In the case that you simply want a connection to the P2P network (ie in the vast majority of cases), all you need to do is instantiate a PeerGroup and connect a few objects:
After this code completes you’ve connected to some peers and fully downloaded the blockchain up to the latest block, filling in missing wallet transactions as it goes. peerGroup.startAndWait() and peerGroup.downloadBlockChain() can be replaced with asynchronous versions peerGroup.start() followed by peerGroup.startBlockchainDownload(listener) when the future returned by start() completes.
Proxying
If you wish to connect to the network using a SOCKS proxy, you must use blocking sockets instead of nio ones. This makes network slightly less efficient, but it should not be noticeable for anything short of very heavy workloads. Then you simply set the Java system properties for a SOCKS proxy (as below) and connections will automatically flow over the proxy.
Using the lower level API’s
You can build a Peer at a lower level, controlling the socket to be used, using code like this:
Of course you would provide your own SocketFactory
instead of using the default.
If you want access to a raw stream of messages with no higher level logic, subclass PeerSocketHandler
and override the processMessage
abstract method. This class implements StreamParser
which breaks raw byte streams into the right subclass of Message
for you, and then lets you handle those messages as you see fit. Create instances of your new object and pass them to an implementation of ClientConnectionManager
, typically either BlockingClientManager
or NioClientManager
to use epoll/select based async IO.