This document describes version 37.0 of the protocol used by Xapian's remote backend. The major protocol version increased to 37 in Xapian 1.3.1.
Clients and servers must support matching major protocol versions and the client's minor protocol version must be the same or lower. This means that for a minor protocol version change, you can upgrade first servers and then clients and everything should work during the upgrades.
The protocol assumes a reliable two-way connection across which arbitrary data can be sent - this could be provided by a TCP socket for example (as it is with xapian-tcpsrv), but any such connection could be used. For example, you could used xapian-progsrv across an ssh connection, or even a custom server across a suitable serial connection.
All messages start with a single byte identifying code. A message from client to server has a MSG_XXX identifying code, while a message from server to client has a REPLY_XXX identifying code (but note that a reply might not actually be in response to a message - REPLY_UPDATE is sent by the server when the connection is opened - and some messages result in multiple replies).
The identifying code is followed by the encoded length of the contents followed by the contents themselves.
Inside the contents, strings are generally passed as an encoded length followed by the string data (this is indicated below by L<...>) except when the string is the last or only thing in the contents in which case we know the length because we know the length of the contents so we don't need to explicitly specify it (indicated by <...> below).
Integers are encoded using the same encoding used for string lengths (indicated by I<...> below).
Floating pointing values are passed using a bit packed encoding of the sign and exponent and a base-256 encoding of the mantissa which avoids any rounding issues (assuming that both machines have FLT_RADIX equal to some power of 2). This is indicated by F<...> below.
Boolean values are passed as a single byte which is the ASCII character value for 0 or 1. This is indicated by B<...> below.
Unsigned byte values are indicated by C<...> below.
The protocol major and minor versions are passed as a single byte each (e.g. '\x1e\x01' for version 30.1). The server and client must understand the same protocol major version, and the server protocol minor version must be greater than or equal to that of the client (this means that the server understands newer MSG_XXX, but will only send newer REPLY_YYY in response to an appropriate client message.
If an unknown exception is caught by the server, this message is sent but with empty contents.
This message can be sent at any point - the serialised exception is unserialised by the client and thrown. The server and client both abort any current sequence of messages.
The reply message is the same format as the server's opening greeting given above.
If write access isn't supported or the database is locked by another writer, then an exception is thrown.
By default the server is also read-only, even if writing is supported. If the client wants to be able to write, it needs to request this explicitly. We do this so that the same server can support multiple read-only clients and one writing client at once, without the protocol for read-only clients requiring an extra message. The overhead of an extra message exchange for a writer is unlikely to matter as indexing is rarely so real-time critical as searching.
If the database was already at the latest version, REPLY_DONE is returned.
If it was reopened, then the reply message is the same format as the server's opening greeting given above.
docid order is '0', '1' or '2'.
sort by is '0', '1', '2' or '3'.
Since positions must be strictly monotonically increasing, we encode (pos - lastpos - 1) so that small differences between large position values can still be encoded compactly. The first position is encoded as its true value.
Since document IDs in postlists must be strictly monotonically increasing, we encode (docid - lastdocid - 1) so that small differences between large document IDs can still be encoded compactly. The first document ID is encoded as its true value - 1 (since document IDs are always > 0).
No reply is sent - this message signals that the client has ended the session.
Only useful for a WritableDatabase (since the same statistics are sent when the connection is initiated in the REPLY_GREETING and they don't change if the database can't change).