18.1. Possible causes of losing log messages

During the course of a message from the sending application to the final destination of the message, there are a number of locations where a message may be lost, even though syslog-ng does its best to avoid message loss. Usually losing messages can be avoided with careful planning and proper configuration of syslog-ng and the hosts running syslog-ng. The following list shows the possible locations where messages may be lost, and provides methods to minimize the risk of losing messages.

Note

The following list covers the main possibilities of losing messages, but does not take into account the possible use of flow-control (for details, see Section 8.2, Managing incoming and outgoing messages with flow-control). This topic will be addressed in more detail in the future releases of this guide.

  • Between the application and the syslog-ng client: Make sure to use an appropriate source to receive the logs from the application (for example from /dev/log). For example, use unix-stream instead of unix-dgram whenever possible.

  • When syslog-ng is sending messages: If syslog-ng cannot send messages to the destination and the output buffer gets full, syslog-ng will drop messages. Use flags (flow-control) to avoid it (for details, see Section 8.2.2, Configuring flow-control). The number of dropped messages is displayed per destination in the log message statistics of syslog-ng (for details, see Chapter 16, Statistics of syslog-ng).

  • On the network: When transferring messages using the UDP protocol, messages may be lost without any notice or feedback — such is the nature of the UDP protocol. Always use the TCP protocol to transfer messages over the network whenever possible.

  • In the socket receive buffer: When transferring messages using the UDP protocol, the UDP datagram (that is, the message) that reaches the receiving host placed in a memory area called the socket receive buffer. If the host receives more messages than it can process, this area overflows, and the kernel drops messages without letting syslog-ng know about it. Using TCP instead of UDP prevents this issue. If you must use the UDP protocol, increase the size of the receive buffer using the so-rcvbuf() option.

  • When syslog-ng is receiving messages:

    • The receiving syslog-ng (for example the syslog-ng server or relay) may drop messages if the fifo of the destination file gets full. The number of dropped messages is displayed per destination in the log message statistics of syslog-ng (for details, see Chapter 16, Statistics of syslog-ng).

  • When the destination cannot handle large load: When syslog-ng is sending messages at a high rate into an SQL database, a file, or another destination, it is possible that the destination cannot handle the load, and processes the messages slowly. As a result, the buffers of syslog-ng fill up, syslog-ng cannot process the incoming messages, and starts to loose messages. For details, see the previous entry. Use the throttle parameter to avoid this problem.

  • As a result of an unclean shutdown of the syslog-ng server: If the host running the syslog-ng server experiences an unclean shutdown, it takes time until the clients realize that the connection to the syslog-ng server is down. Messages that are put into the output TCP buffer of the clients during this period are not sent to the server.

  • When syslog-ng OSE is writing messages into files: If syslog-ng OSE receives a signal (SIG) while writing log messages to file, the log message that is processed by the write call can be lost if the flush_lines parameter is higher than 1.