pytweezer.servers.sipyco_patches module¶
Runtime patches applied to the vendored sipyco RPC library.
Imported for its side effects by pytweezer.servers, so every pytweezer
process that talks to a device server gets them. apply_patches() is
idempotent and safe to call again.
Reply reading¶
sipyco.pc_rpc.Client reads each reply with _socket_readline, which
accumulates 4 kB chunks into a str with buf += chunk while testing
buf.find("\n", offset) in the while condition. That shape stops CPython
from appending in place, so every chunk copies the whole reply built so far and
the read costs O(n^2) in the reply size. Camera frames are the payloads big
enough for it to hurt: a 100-frame 512x512 uint16 grab (52 MB) takes ~200 s to
read on loopback, against ~0.7 s here. The replacement accumulates into a list
and joins once, and reads 64 kB per recv instead of 4 kB.
Only the synchronous Client is affected. AsyncioClient already reads
through asyncio’s C-implemented readline with a 100 MB line limit.
- pytweezer.servers.sipyco_patches.RECV_BUFSIZE = 65536¶
recvsize for reply reading. Large replies are camera frames of a few MB to tens of MB, so the 4 kB sipyco default costs thousands of extra syscalls.