Skip to content

Latest commit

 

History

History
176 lines (130 loc) · 8.49 KB

README.md

File metadata and controls

176 lines (130 loc) · 8.49 KB

zmqpush

Asynchronously queue STDIN and push across the wire using a ZeroMQ socket.

Built with the new (in python3.4) asyncio module and aiozmq.

Utilizes NONBLOCK-ing stdin and Edge-Triggered epoll() to detect content in stdin.

Designed to recieve input via a unix pipe and send to Logstash zeromq input.

Currently, it formats every line from stdin into the following JSON:

jsonmsg = '{"message":"%s","type":"%s","@pid":%d}' % (quote_escape(line),
                                                                   logtype,
                                                                   pid)

zmqpush will take the 1st command-line argument and assign it to logtype.

If no argument is specified, logtype is set to "syslog" in order to compensate for rsyslog OMProg's inability to pass arguments.

logtype variable is put into the JSON as "type", which allows it to be automatically parsed and utilized directly by Logstash.

@pid is the currently executing zmqpush process ID, and was added mostly for debugging, however it remains since it may be useful down the line.

Configuration

## Configuration variables
loghost = "logstash.host"
logport = "5014"
##

Examples

Shell pipes:

$ echo test | zmqpush.py
Processed 1 messages in 2.2003ms. Tagged with @pid:19867

$ seq 1 100 | zmqpush.py
Processed 100 messages in 37.0543ms. Tagged with @pid:19863

$ for x in {1..3}; do seq 1 100 | zmqpush.py && sleep 1; done
Processed 100 messages in 52.6803ms. Tagged with @pid:19848
Processed 100 messages in 48.8532ms. Tagged with @pid:19853
Processed 100 messages in 50.1794ms. Tagged with @pid:19858

$ while true; do seq 1 100 && sleep 1; done | zmqpush.py 
^CProcessed 300 messages in 2460.1966ms. Tagged with @pid:17301

rsyslog5: Ship all messages with RFC5424 format

$ModLoad            omprog
$ActionOMProgBinary /path/to/zmqpush.py
*.*                 :omprog:;RSYSLOG_SyslogProtocol23Format

Details

zmqpush is essentially made up of two functions, decorated as asyncio.coroutines. They are launched in parallel, and in order to prevent a race condition on startup we utilize asyncio.Future objects to coordinate between the coroutines.

zmq_pusher() will start the ZeroMQ socket, update ZMQFuture object, then enter a while True loop in which it perpetually writes messages from the queue to the ZeroMQ socket.

stdin_queuer() will wait for the ZMQFuture object to be updated, then enter a while True loop in which it perpetually polls for input on stdin, which it places into the queue.

If input is detected during the poll cycle, the inner for line in sys.stdin loop is initiated to pull all available input into the queue. If there is a continuous stream of input, stdin_queuer() will most likely not leave the inner for loop, therefor a yield is given after every line from stdin is queued so zmq_pusher() can pick up the message and immediately write it to the ZeroMQ socket. This allows continuous streams of input to be shipped via zmq_pusher() in a practically synchronous manner.

If no input is detected, stdin_queuer() just yields, which allows zmq_pusher() the chance to clear the ZeroMQ socket buffer and/or queue, if nescessary. If there haven't been any connectivity issues, zmq_pusher() will have nothing to do, so it instantly comes back to stdin_queuer() and the poll cycle starts again, repeat ad infinitum.

50ms for the poll cycle (timeout) could be unnescessarily frequent, however further down the load is shown to be minimal on idle, which is the only scenario that will show any really difference between poll timeout settings. Having it this low also allows zmq_pusher() to both quickly ship incoming input and recover after connectivity issues.

If there are connectivity issues, zmq_pusher() has logic to catch when the ZeroMQ socket buffer has exceeded the low watermark (described below), which will stop it from getting any further messages from the queue or writing them to the socket. Instead zmq_pusher() will initiate an asynchronous buffer drain on the socket, then initiate an asynchronous sleep for 100ms - which will yield back to stdin_queuer(). Hence, input will continue to be placed into the queue, which will grow in application memory. Depending on the amount of input, and how long connectivity is down, the memory footprint could grow minimally or massively - which may be a possible hazard in some extreme situation. Queue size limit could be configured so there's some sort of ceiling, however it is currently unlimited.

The reasoning behind all this is to provide resilient log reporting from the source by queueing messages and ensuring there is no buffer overflow during connectivity outages. We also compensate for rsyslog5's synchronous processing - messages sent through syslog will never hang on the OMProg output given that we set sys.stdin nonblocking. When using the default blocking sys.stdin file descriptor, rsyslog OMProg's stdout pipe could fill up and then start blocking, which stopped rsyslog processing and prevented messages from being sent to the local file destinations until the OMProg->zmqpush pipe was cleared. (BAD!)

A possible improvement to this application would be to move the queue to disk periodically, maybe during the buffer drain cycles, then read from disk during recovery when connection is back up.

The buffer watermarks are configurable, however by default ZMQ automatically sets them based within the OS socket settings, which is probably ideal. Since we're writing to a PUSH socket, let's compare against TCP write memory:

# ./zmqpush-watermark.py 
Low watermark: 16384
High watermark: 65536

# sysctl -a | grep 'tcp_wmem'
net.ipv4.tcp_wmem = 4096	16384	4194304

Stats

Following tests were performed on a single-core VM, with zmqpush running via rsyslog, comparing various poll timeouts-

No load:

$ pidstat -C zmqpush.py 1 60 -u -w
..

50ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        25873    0.50    0.02    0.00    0.52     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        25873     19.90      0.20  zmqpush.py

75ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        25931    0.37    0.02    0.00    0.38     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        25931     13.30      0.10  zmqpush.py

100ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        25965    0.28    0.02    0.00    0.30     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        25965      9.98      0.07  zmqpush.py

Slight load:

$ while true; do seq 1 100 | logger -t test && sleep 1; done &
$ pidstat -C zmqpush.py 1 60 -u -w
..

50ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        27854    2.14    0.64    0.00    2.78     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        27854     25.28     97.36  zmqpush.py

75ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        27515    2.06    0.62    0.00    2.68     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        27515     19.75     97.04  zmqpush.py

100ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        25965    1.99    0.55    0.00    2.54     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        25965     15.21     97.69  zmqpush.py

High load:

$ while true; do seq 1 100 | logger -t test && sleep 0.2; done &
$ pidstat -C zmqpush.py 1 60 -u -w
..

50ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        27854    9.00    4.53    0.00   13.53     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        27854     44.29    456.13  zmqpush.py

75ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        29581    8.83    4.62    0.00   13.44     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        29581     39.09    456.25  zmqpush.py

100ms:
Average:          PID    %usr %system  %guest    %CPU   CPU  Command
Average:        30805    9.04    4.55    0.00   13.59     -  zmqpush.py
Average:          PID   cswch/s nvcswch/s  Command
Average:        30805     33.75    457.31  zmqpush.py