diff options
Diffstat (limited to 'lib/kernel/doc/src/logger_chapter.xml')
| -rw-r--r-- | lib/kernel/doc/src/logger_chapter.xml | 178 |
1 files changed, 176 insertions, 2 deletions
diff --git a/lib/kernel/doc/src/logger_chapter.xml b/lib/kernel/doc/src/logger_chapter.xml index 2a325453da..0bc3b37476 100644 --- a/lib/kernel/doc/src/logger_chapter.xml +++ b/lib/kernel/doc/src/logger_chapter.xml @@ -296,6 +296,7 @@ </section> <section> + <marker id="handler_configuration"/> <title>Handler configuration</title> <taglist> <tag><c>level</c></tag> @@ -577,8 +578,8 @@ log(#{msg:={F,A}},#{myhandler_fd:=Fd}) -> <p>For examples of overload protection, please refer to the implementation of <seealso marker="logger_std_h"><c>logger_std_h</c></seealso> - and <!--<seealso marker="logger_disk_log_h"--><c>logger_disk_log_h</c> - <!--/seealso-->.</p> + and <seealso marker="logger_disk_log_h"><c>logger_disk_log_h</c> + </seealso>.</p> <p>Below is a simpler example of a handler which logs through one single process, and uses the default formatter to gain a common @@ -632,6 +633,179 @@ my_report_cb(R) -> </code> </section> + <section> + <marker id="overload_protection"/> + <title>Protecting the handler from overload</title> + <p>In order for the built-in handlers to survive, and stay responsive, + during periods of high load (i.e. when huge numbers of incoming + log requests must be handled), a mechanism for overload protection + has been implemented in the + <seealso marker="logger_std_h"><c>logger_std_h</c></seealso> + and <seealso marker="logger_disk_log_h"><c>logger_disk_log_h</c> + </seealso> handler. The mechanism, used by both handlers, works + as follows:</p> + + <section> + <title>Message queue length</title> + <p>The handler process keeps track of the length of its message + queue and reacts in different ways depending on the current status. + The purpose is to keep the handler in, or (as quickly as possible), + get the handler into, a state where it can keep up with the pace + of incoming log requests. The memory usage of the handler must never + keep growing larger and larger, since that would eventually cause the + handler to crash. Three thresholds with associated actions have been + defined:</p> + + <taglist> + <tag><c>toggle_sync_qlen</c></tag> + <item> + <p>The default value of this level is <c>10</c> messages, + and as long as the length of the message queue is lower, all log + requests are handled asynchronously. This simply means that the + process sending the log request (by calling a log function in the + logger API) does not wait for a response from the handler but + continues executing immediately after the request (i.e. it will not + be affected by the time it takes the handler to print to the log + device). If the message queue grows larger than this value, however, + the handler starts handling the log requests synchronously instead, + meaning the process sending the request will have to wait for a + response. When the handler manages to reduce the message queue to a + level below the <c>toggle_sync_qlen</c> threshold, asynchronous + operation is resumed. The switch from asynchronous to synchronous + mode will force the logging tempo of few busy senders to slow down, + but can not protect the handler sufficiently in situations of many + concurrent senders.</p> + </item> + <tag><c>drop_new_reqs_qlen</c></tag> + <item> + <p>When the message queue has grown larger than this threshold, which + defaults to <c>200</c> messages, the handler switches to a mode in + which it drops any new requests being made. Dropping a message in + this state means that the log function never actually sends a message + to the handler. The log call simply returns without an action. When + the length of the message queue has been reduced to a level below this + threshold, synchronous or asynchronous request handling mode is + resumed.</p> + </item> + <tag><c>flush_reqs_qlen</c></tag> + <item> + <p>Above this threshold, which defaults to <c>1000</c> messages, a + flush operation takes place, in which all messages buffered in the + process mailbox get deleted without any logging actually taking + place. (Processes waiting for a response from a synchronous log request + will receive a reply indicating that the request has been dropped).</p> + </item> + </taglist> + + <p>For the overload protection algorithm to work properly, it is a + requirement that:</p> + + <p><c>toggle_sync_qlen < drop_new_reqs_qlen < flush_reqs_qlen</c></p> + + <p>During high load scenarios, the length of the handler message queue + rarely grows in a linear and predictable way. Instead, whenever the + handler process gets scheduled in, it can have an almost arbitrary number + of messages waiting in the mailbox. It's for this reason that the overload + protection mechanism is focused on acting quickly and quite drastically + (such as immediately dropping or flushing messages) as soon as a large + queue length is detected. </p> + + <p>The thresholds listed above may be modified by the user if, e.g, a handler + shouldn't drop or flush messages unless the message queue length grows + extremely large. (The handler must be allowed to use large amounts of memory + under such circumstances however). Another example of when the user might want + to change the settings is if, for performance reasons, the logging processes must + never get blocked by synchronous log requests, while dropping or flushing requests + is perfectly acceptable (since it doesn't affect the performance of the + loggers).</p> + + <p>A configuration example:</p> + <code type="none"> +logger:add_handler(my_standard_h, logger_std_h, + #{logger_std_h => + #{type => {file,"./system_info.log"}, + toggle_sync_qlen => 100, + drop_new_reqs_qlen => 1000, + flush_reqs_qlen => 2000}}). + </code> + </section> + + <section> + <title>Controlling bursts of log requests</title> + <p>A potential problem with large bursts of log requests, is that log files + may get full or wrapped too quickly (in the latter case overwriting + previously logged data that could be of great importance). For this reason, + both built-in handlers offer the possibility to set a maximum level of how + many requests to process with a certain time frame. With this burst control + feature enabled, the handler will take care of bursts of log requests + without choking log files, or the console, with massive amounts of + printouts. These are the configuration parameters:</p> + + <taglist> + <tag><c>enable_burst_limit</c></tag> + <item> + <p>This is set to <c>true</c> by default. The value <c>false</c> + disables the burst control feature.</p> + </item> + <tag><c>burst_limit_size</c></tag> + <item> + <p>This is how many requests should be processed within the + <c>burst_window_time</c> time frame. After this maximum has been + reached, successive requests will be dropped until the end of the + time frame. The default value is <c>500</c> messages.</p> + </item> + <tag><c>burst_window_time</c></tag> + <item> + <p>The default window is <c>1000</c> milliseconds long.</p> + </item> + </taglist> + + <p>A configuration example:</p> + <code type="none"> +logger:add_handler(my_disk_log_h, logger_disk_log_h, + #{disk_log_opts => + #{file => "./my_disk_log"}, + logger_disk_log_h => + #{burst_limit_size => 10, + burst_window_time => 500}}). + </code> + </section> + + <section> + <title>Terminating a large handler</title> + <p>A handler process may grow large even if it can manage peaks of high load + without crashing. The overload protection mechanism includes user configurable + levels for a maximum allowed message queue length and maximum allowed memory + usage. This feature is disabled by default, but can be switched on by means + of the following configuration parameters:</p> + + <taglist> + <tag><c>enable_kill_overloaded</c></tag> + <item> + <p>This is set to <c>false</c> by default. The value <c>true</c> + enables the feature.</p> + </item> + <tag><c>handler_overloaded_qlen</c></tag> + <item> + <p>This is the maximum allowed queue length. If the mailbox grows larger + than this, the handler process gets terminated.</p> + </item> + <tag><c>handler_overloaded_mem</c></tag> + <item> + <p>This is the maximum allowed memory usage of the handler process. If + the handler grows any larger, the process gets terminated.</p> + </item> + <tag><c>handler_restart_after</c></tag> + <item> + <p>If the handler gets terminated because of its queue length or + memory usage, it can get automatically restarted again after a + configurable delay time. The time is specified in milliseconds + and <c>5000</c> is the default value. The value <c>never</c> can + also be set, which prevents a restart.</p> + </item> + </taglist> + </section> + </section> <section> <title>See Also</title> |