mtail Language Reference


As mtail is designed to tail log files and apply regular expressions to new log lines to extract data, the language naturally follows this pattern-action style.

It resembles another, more famous pattern-action language, that of AWK.

This page errs on the side of a language specification and reference. See the Programming Guide for a gentler introduction to writing mtail programs.

Program Execution

mtail runs all programs on every line received by the log tailing subsystem. The rough model of this looks like:

for line in lines:
  for regex in regexes:
    if match:
      do something

Each program operates once on a single line of log data, and then terminates.

Program Structure

An mtail program consists of exported variable definitions, pattern-action statements, and optional decorator definitions.

exported variable

pattern {
  action statements

def decorator {
  pattern and action statements

Exported Variables

mtail’s purpose is to extract information from logs and deliver them to a monitoring system. Thus, variables must be named for export.

Variables, which have type counter or gauge, must be declared before their use.

counter lines_total
gauge queue_length

They can be exported with a different name, with the as keyword, if one wants to use characters that would cause a parse error. This example causes the metric to be named line-count in the collecting monitoring system.

counter lines_total as "line-count"

Variables can be dimensioned with one or more axes, with the by keyword, creating multidimensional data. Dimensions can be used for creating histograms, as well.

counter bytes by operation, direction
counter latency_ms by bucket

Putting the hidden keyword at the start of the declaration means it won’t be exported, which can be useful for storing temporary information. This is the only way to share state between each line being processed.

hidden counter login_failures

Pattern/Action form.

mtail programs look a lot like awk programs. They consist of a conditional expression followed by a brace-enclosed block of code:


COND is a conditional expression. It can be a regular expression, which if matched enters the action block, or a relational expression as you might encounter in a C program’s if statement (but without the if, it is implicit.)

/foo/ {

variable > 0 {

/foo/ && variable > 0 {

In the above program, ACTION1 is taken on each line input if that line matches the word foo, and ACTION2 is taken on each line if when that line is read, the variable variable is greater than 0. ACTION3 occurs if both are true.

The action statements must be wrapped in curly braces, i.e. {}. mtail programs have no single-line statement conditionals like C.

Regular Expressions

mtail supports RE2-style regular expression syntax, but is limited by what is supported by the Go implementation of Go’s regexp/syntax.

Constant pattern fragments

To re-use parts of regular expressions, you can assign them to a const identifier:

const PREFIX /^\w+\W+\d+ /


PREFIX + /foo/ {

In this example, ACTION1 is done for every line that starts with the prefix regex, and ACTION2 is done for the subset of those lines that also contain ‘foo’.

Pattern fragments like this don’t need to be prefixes, they can be anywhere in the expression.

counter maybe_ipv4

const IPv4 /(?P<ip>\d+\.\d+\.\d+\.\d+)/

/something with an / + IPv4 + / address/ {

See dhcpd.mtail for more examples of this.

See also the section on decorators below for improving readability of expressions that are only matched once.


More complex expressions can be built up from relational expressions and other pattern expressions.


The following relational operators are available in mtail:

The following arithmetic operators are available in mtail:

The following arithmetic operators act on exported variables.

else Clauses

When a conditional expression does not match, action can be taken as well:

/foo/ {
} else {

Else clauses can be nested. There is no ambiguity with the dangling-else problem, as mtail programs must wrap all block statements in {}.

otherwise clauses

The otherwise keyword can be used as a conditional statement. It matches if no preceding conditional in the current scope has matched. This behaves similarly to the default clause in a C switch statement.

/foo/ {
  /foo1/ {
  /foo2/ {
  otherwise {

In this example, ACTION3 will be executed if neither /foo1/ or /foo2/ match on the input, but /foo/ does.


Incrementing a Counter

The simplest mtail program merely counts lines read:

/$/ {

This program instructs mtail to increment the lines_total counter variable on every line received (specifically anytime an end-of-line is matched.)

Capture Groups

Regular expressions in patterns can contain capture groups – subexpressions wrapped in parentheses. These can be referred to in the action block to extract data from the line being matched.

For example, part of a program that can extract from rsyncd logs may want to break down transfers by operation and module.

counter transfers_total by operation, module

/(?P<operation>\S+) (\S+) \[\S+\] (\S+) \(\S*\) \S+ (?P<bytes>\d+)/ {

Or, the value of the counter can be increased by the value of a capture group:

counter bytes_total by operation, module

/(?P<operation>\S+) (\S+) \[\S+\] (\S+) \(\S*\) \S+ (?P<bytes>\d+)/ {
  bytes_total[$operation][$3] += $bytes

Numeric capture groups address subexpressions in the match result as you might expect from regular expression groups in other languages, like awk and perl – e.g. the expression $3 refers to the third capture group in the regular expression.

Named capture groups can be referred to by their name as indicated in the regular expression using the ?P<name> notation, as popularised by the Python regular expression library – e.g. $bytes refers to (?P<bytes>\d+) in the examples above.

Capture groups can be used in the same expression that defines them, for example in this expression that matches and produces $x, then compares against that value.

/(?P<x>\d+)/ && $x > 1 {


It is also useful to timestamp a metric with the time the application thought an event occurred. Logs typically prefix the log line with a timestamp string, which can be extracted and then parsed into a timestamp internally, with the strptime builtin function.

A regular expression that extracts the timestamp in boring old syslog format looks like:

/^(?P<date>\w+\s+\d+\s+\d+:\d+:\d+)/ {
  strptime($date, "Jan 02 15:04:05")

Buyer beware! The format string used by mtail is the same as the Go time.Parse() format string, which is completely unlike that used by C’s strptime. The format string must always be the 2nd of January 2006 at 3:04:05 PM. See the documentation for the ANSIC format in the above link for more details. NOTE that unlike Go’s time.Parse() (and like C’s) the format string is the second argument to this builtin function.

NOTE: without a strptime() call, mtail will default to using the current system time for the timestamp of the event. This may be satisfactory for near-real-time logging.

Nested Actions

It is of course possible to nest more pattern-actions within actions. This lets you factor out common parts of a match expression and deal with per-message actions separately.

For example, parsing syslog timestamps is something you may only wish to do once, as it’s expensive to match (and difficult to read!)

counter foo
counter bar

/^(?P<date>\w+\s+\d+\s+\d+:\d+:\d+)/ {
  strptime($date, "Jan 02 15:04:05")

  /foo/ {

  /bar/ {

This will result in both foo and bar counters being timestamped with the current log line’s parsed time, once they match a line.

Decorated actions

Decorated actions are an inversion of nested actions. They allow the program to define repetitive functions that perform the same extraction across many different actions.

For example, most log file formats start with a timestamp prefix. To reduce duplication of work, decorators can be used to factor out the common work of extracting the timestamp. For example, to define a decorator, use the def keyword:

def syslog {
  /(?P<date>\w+\s+\d+\s+\d+:\d+:\d+)/ {
    strptime($date, "Jan  2 15:04:05")

The decorator definition starts and ends in a curly-braced block, and looks like a normal pattern/action as above. The new part is the next keyword, which indicates to mtail where to jump into the decorated block.

To use a decorator:

@syslog {
  /some event/ {

The @ notation, familiar to Python programmers, denotes that this block is “wrapped” by the syslog decorator. The syslog decorator will be called on each line first, which extracts the timestamp of the log line. Then, next causes the wrapped block to execute, so then mtail matches the line against the pattern some event, and if it does match, increments variable.


mtail metrics have a kind and a type. The kind affects how the metric is recorded, and the type describes the data being recorded.

Ordinarily mtail doesn’t treat kinds specially, except when they are being exported.

The second dimension is the internal representation of a value, which is used by mtail to attempt to generate efficient bytecode.

Some of these types can only be used in certain locations – for example, you can’t increment a counter by a string, but mtail will fall back to a attempt to do so, logging an error if a runtime type conversion fails. Likewise, the only type that a histogram can observe is a Float.

These types are usually inferred from use, but can be influenced by the programmer with builtin functions. Read on.

Builtin functions

mtail contains some builtin functions for help with extracting information and manipulating state.

There are “pure” builtin functions, in that they have no side effects on the program state.

There are type coercion functions, useful for overriding the type inference made by the compiler if it chooses badly. (If the choice is egregious, please file a bug!)

A few builtin functions exist for manipulating the virtual machine state as side effects for the metric export.

The current timestamp register refers to mtail’s idea of the time associated with the current log line. This timestamp is used when the variables are exported to the upstream collector. The value defaults to the time that the log line arrives in mtail, and can be changed with the settime() or strptime() builtins.

User defined functions are not supported, but read on to Decorated Actions for how to reuse common code.

Numerical capture groups and Metric type information

By limiting the pattern of a capturing group to only numeric characters, the programmer can provide hints to mtail about the type of an expression. For example, in the regular expression


the first capture group can only match digits, and so the compiler will infer that this is an integer match.


looks like it matches floating point numbers, and so the compiler will infer that this is of type float.

NOTE: In the expression above, the dot is escaped. A regular expression operator . matches every character and so the inference assumes that the type of ‘.’ is a string.

The compiler performs type inference on the expressions that use the capture groups, and the metrics they are ultimately assigned to, and will assign a type (either integer or floating point) to the metrics exported.

Thus in a program like:

gauge i
gauge f

/(\d+)/ {
  i = $1

/(\d+\.\d+)/ {
  f = $1

the metric i will be of type Int and the metric f will be of type Float.

The advantage of limiting pattern matches to specific values is that mtail can generate faster bytecode if it knows at compile-time the types to expect. If mtail can’t infer the value types, they default to String and mtail will attempt a value conversion at runtime if necessary. Runtime conversion errors will be emitted to the standard INFO log, and terminate program execution for that log line.

Variable Storage Management


mtail performs no implicit garbage collection in the metric storage. The program can hint to the virtual machine that a specific datum in a dimensioned metric is no longer going to be used with the del keyword.

gauge duration by session
hidden session_start by session

/end/ {
  duration[$session] = timestamp() - session_start[$session]

  del session_start[$session]

In this example, a hidden metric is used to record some internal state. It will grow unbounded as the number of sessions increases. If the programmer knows that the /end/ pattern is the last time a session will be observed, then the datum at $session will be freed, which keeps mtail memory usage under control and will improve search time for finding dimensioned metrics.

del can be modified with the after keyword, signalling that the metric should be deleted after some period of no activity. For example, the expression

  del session_start[$session] after 24h

would mean that the datum indexed by $session will be removed 24 hours after the last update is recorded.

The del-after form takes any time period supported by the go time.ParseDuration function.

Expiry is only processed once ever hour, so durations shorter than 1h won’t take effect until the next hour has passed.

This command only makes sense for dimensioned metrics.


A size limit can be specified on a metric with the modifier limit.

counter bytes_total by operation limit 500

When the garbage collection run encounters a variable with size limit that is over its size limit, it will remove the oldest values until the whole metric is below its limit again. Oldest values are chosen by the timestamp of the datum.

This modifier only makes sense for dimensioned metrics.

Stopping the program

The program runs from start to finish once per line, but sometimes you may want to stop the program early. For example, if the log filename does not match a pattern, or some stateful metric indicates work shouldn’t be done.

For this purpose, the stop keyword terminates the program immediately.

The simplest and most useless mtail program is thus:


But for more useful situations, perhaps stopping if the log filename doesn’t match a pattern:

getfilename() !~ /apache.access.log/ {