Whereas the Unix toolkit is made of the separate executables cat, tail, cut,
sort, etc., Miller has subcommands, invoked as follows:
POKI_INCLUDE_ESCAPED(data/subcommand-example.txt)HERE
These draw from other sources (see also POKI_PUT_LINK_FOR_PAGE(originality.html)HERE):
count-distinct is SQL-ish, and
rename can be done by sed (which does it faster:
see POKI_PUT_LINK_FOR_PAGE(performance.html)HERE).
These are as discussed in POKI_PUT_LINK_FOR_PAGE(file-formats.html)HERE, with the exception of --right
which makes pretty-printed output right-aligned:
Use --csv, --pprint, etc. when the input and output formats are the same.
Use --icsv --opprint, etc. when you want format conversion as part of what Miller does to your data.
DKVP (key-value-pair) format is the default for input and output. So,
--oxtab is the same as --idkvp --oxtab.
In-place mode
Use the mlr -I flag to process files in-place. For example,
mlr -I --csv cut -x -f unwanted_column_name mydata/*.csv will remove
unwanted_column_name from all your *.csv files in your
mydata/ subdirectory.
By default, Miller output goes to the screen (or you can redirect a file
using > or to another process using |). With -I,
for each file name on the command line, output is written to a temporary file
in the same directory. Miller writes its output into that temp file, which is
then renamed over the original. Then, processing continues on the next file.
Each file is processed in isolation: if the output format is CSV, CSV headers
will be present in each output file; statistics are only over each file's own
records; and so on.
Please see here
for examples.
Compression
Options:
--prepipe {command}
The prepipe command is anything which reads from standard input and produces data acceptable to
Miller. Nominally this allows you to use whichever decompression utilities you have installed on your
system, on a per-file basis. If the command has flags, quote them: e.g. mlr --prepipe 'zcat -cf'. Examples:
# These two produce the same output:
$ gunzip < myfile1.csv.gz | mlr cut -f hostname,uptime
$ mlr --prepipe gunzip cut -f hostname,uptime myfile1.csv.gz
# With multiple input files you need --prepipe:
$ mlr --prepipe gunzip cut -f hostname,uptime myfile1.csv.gz myfile2.csv.gz
$ mlr --prepipe gunzip --idkvp --oxtab cut -f hostname,uptime myfile1.dat.gz myfile2.dat.gz
# Similar to the above, but with compressed output as well as input:
$ gunzip < myfile1.csv.gz | mlr cut -f hostname,uptime | gzip > outfile.csv.gz
$ mlr --prepipe gunzip cut -f hostname,uptime myfile1.csv.gz | gzip > outfile.csv.gz
$ mlr --prepipe gunzip cut -f hostname,uptime myfile1.csv.gz myfile2.csv.gz | gzip > outfile.csv.gz
# Similar to the above, but with different compression tools for input and output:
$ gunzip < myfile1.csv.gz | mlr cut -f hostname,uptime | xz -z > outfile.csv.xz
$ xz -cd < myfile1.csv.xz | mlr cut -f hostname,uptime | gzip > outfile.csv.xz
$ mlr --prepipe 'xz -cd' cut -f hostname,uptime myfile1.csv.xz myfile2.csv.xz | xz -z > outfile.csv.xz
... etc.
Record/field/pair separators
Miller has record separators IRS and ORS, field
separators IFS and OFS, and pair separators IPS and
OPS. For example, in the DKVP line a=1,b=2,c=3, the record
separator is newline, field separator is comma, and pair separator is the
equals sign. These are the default values.
Options:
You can change a separator from input to output via e.g. --ifs =
--ofs :. Or, you can specify that the same separator is to be used for
input and output via e.g. --fs :.
The pair separator is only relevant to DKVP format.
Pretty-print and xtab formats ignore the separator arguments altogether.
The --repifs means that multiple successive occurrences of the
field separator count as one. For example, in CSV data we often signify nulls
by empty strings, e.g. 2,9,,,,,6,5,4. On the other hand, if the field
separator is a space, it might be more natural to parse 2 4 5 the
same as 2 4 5: --repifs --ifs ' ' lets this happen. In fact,
the --ipprint option above is internally implemented in terms of
--repifs.
Just write out the desired separator, e.g. --ofs '|'. But you
may use the symbolic names newline, space, tab,
pipe, or semicolon if you like.
Number formatting
The command-line option --ofmt {format string} is the global
number format for commands which generate numeric output, e.g.
stats1, stats2, histogram, and step, as
well as mlr put. Examples:
POKI_CARDIFY(--ofmt %.9le --ofmt %.6lf --ofmt %.0lf)HERE
These are just C printf formats applied to double-precision
numbers. Please don’t use %s or %d. Additionally, if
you use leading width (e.g. %18.12lf) then the output will contain
embedded whitespace, which may not be what you want if you pipe the output to
something else, particularly CSV. I use Miller’s pretty-print format
(mlr --opprint) to column-align numerical data.
To apply formatting to a single field, overriding the global
ofmt, use fmtnum function within mlr put. For example:
POKI_RUN_COMMAND{{echo 'x=3.1,y=4.3' | mlr put '$z=fmtnum($x*$y,"%08lf")'}}HERE
POKI_RUN_COMMAND{{echo 'x=0xffff,y=0xff' | mlr put '$z=fmtnum(int($x*$y),"%08llx")'}}HERE
Input conversion from hexadecimal is done automatically on fields handled
by mlr put and mlr filter as long as the field value begins
with "0x". To apply output conversion to hexadecimal on a single column, you
may use fmtnum, or the keystroke-saving hexfmt function.
Example:
POKI_RUN_COMMAND{{echo 'x=0xffff,y=0xff' | mlr put '$z=hexfmt($x*$y)'}}HERE
In accord with the
Unix philosophy, you can pipe data into or out of
Miller. For example:
POKI_CARDIFY(mlr cut --complement -f os_version *.dat | mlr sort -f hostname,uptime)HERE
You can, if you like, instead simply chain commands together using the
then keyword:
POKI_CARDIFY(mlr cut --complement -f os_version then sort -f hostname,uptime *.dat)HERE
(You can precede the very first verb with then, if you like, for symmetry.)
Here’s a performance comparison:
POKI_INCLUDE_ESCAPED(data/then-chaining-performance.txt)HERE
There are two reasons to use then-chaining: one is for performance, although I
don’t expect this to be a win in all cases. Using then-chaining avoids
redundant string-parsing and string-formatting at each pipeline step: instead
input records are parsed once, they are fed through each pipeline stage in
memory, and then output records are formatted once. On the other hand, Miller
is single-threaded, while modern systems are usually multi-processor, and when
streaming-data programs operate through pipes, each one can use a CPU. Rest
assured you get the same results either way.
The other reason to use then-chaining is for simplicity: you don’t
have re-type formatting flags (e.g. --csv --fs tab) at every
pipeline stage.
Auxiliary commands
There are a few nearly-standalone programs which have nothing to do with the rest of Miller, do not
participate in record streams, and do not deal with file formats. They might as well be little standalone executables
but they’re delivered within the main Miller executable for convenience.
POKI_RUN_COMMAND{{mlr aux-list}}HERE
Examples:
POKI_RUN_COMMAND{{echo 'Hello, world!' | mlr hex}}HERE
POKI_RUN_COMMAND{{echo 'Hello, world!' | mlr hex -r}}HERE
POKI_RUN_COMMAND{{echo 'Hello, world!' | mlr hex -r | mlr unhex}}HERE
POKI_RUN_COMMAND{{echo 'Hello, world!' | mlr lecat --mono}}HERE
Data types
Miller’s input and output are all string-oriented: there is (as of
August 2015 anyway) no support for binary record packing. In this sense,
everything is a string in and out of Miller. During processing, field names
are always strings, even if they have names like "3"; field values are usually
strings. Field values’ ability to be interpreted as a non-string type
only has meaning when comparison or function operations are done on them. And
it is an error condition if Miller encounters non-numeric (or otherwise
mistyped) data in a field in which it has been asked to do numeric (or
otherwise type-specific) operations.
Field values are treated as numeric for the following:
Miller’s types for function processing are empty-null (empty
string), absent-null (reads of unset right-hand sides, or fall-through
non-explicit return values from user-defined functions), error,
string, float (double-precision), int (64-bit signed), and
boolean.
On input, string values representable as numbers, e.g. "3" or "3.1", are
treated as int or float, respectively. If a record has x=1,y=2 then
mlr put '$z=$x+$y' will produce x=1,y=2,z=3, and mlr put
'$z=$x.$y' does not give an error simply because the dot operator has been
generalized to stringify non-strings. To coerce back to string for processing,
use the string function: mlr put '$z=string($x).string($y)'
will produce x=1,y=2,z=12.
On input, string values representable as boolean (e.g. "true",
"false") are not automatically treated as boolean. (This is
because "true" and "false" are ordinary words, and auto
string-to-boolean on a column consisting of words would result in some strings
mixed with some booleans.) Use the boolean function to coerce: e.g.
giving the record x=1,y=2,w=false to mlr put '$z=($x<$y) ||
boolean($w)'.
Functions take types as described in mlr --help-all-functions:
for example, log10 takes float input and produces float output,
gmt2sec maps string to int, and sec2gmt maps int to string.
All math functions described in mlr --help-all-functions take
integer as well as float input.
Null data: empty and absent
One of Miller’s key features is its support for heterogeneous
data. For example, take mlr sort: if you try to sort on field
hostname when not all records in the data stream have a field
named hostname, it is not an error (although you could pre-filter the
data stream using mlr having-fields --at-least hostname then sort
...). Rather, records lacking one or more sort keys are simply output
contiguously by mlr sort.
Miller has two kinds of null data:
Empty (key present, value empty): a field name is present in a
record (or in an out-of-stream variable) with empty value: e.g. x=,y=2
in the data input stream, or assignment $x="" or @x="" in
mlr put.
Absent (key not present): a field name is not present, e.g. input
record is x=1,y=2 and a put or filter expression
refers to $z. Or, reading an out-of-stream variable which hasn’t
been assigned a value yet, e.g. mlr put -q '@sum += $x'; end{emit
@sum}' or mlr put -q '@sum[$a][$b] += $x'; end{emit @sum, "a",
"b"}'.
You can test these programatically using the functions
is_empty/is_not_empty, is_absent/is_present, and
is_null/is_not_null. For the last pair, note that null means
either empty or absent.
Rules for null-handling:
Records with one or more empty sort-field values sort after records with
all sort-field values present:
POKI_RUN_COMMAND{{mlr cat data/sort-null.dat}}HERE
POKI_RUN_COMMAND{{mlr sort -n a data/sort-null.dat}}HERE
POKI_RUN_COMMAND{{mlr sort -nr a data/sort-null.dat}}HERE
Functions/operators which have one or more empty arguments produce empty output: e.g.
POKI_RUN_COMMAND{{echo 'x=2,y=3' | mlr put '$a=$x+$y'}}HERE
POKI_RUN_COMMAND{{echo 'x=,y=3' | mlr put '$a=$x+$y'}}HERE
POKI_RUN_COMMAND{{echo 'x=,y=3' | mlr put '$a=log($x);$b=log($y)'}}HERE
with the exception that the min and max functions are
special: if one argument is non-null, it wins:
POKI_RUN_COMMAND{{echo 'x=,y=3' | mlr put '$a=min($x,$y);$b=max($x,$y)'}}HERE
Functions of absent variables (e.g. mlr put '$y =
log10($nonesuch)') evaluate to absent, and arithmetic/bitwise/boolean
operators with both operands being absent evaluate to absent.
Arithmetic operators with one absent operand return the other operand.
More specifically, absent values act like zero for addition/subtraction, and
one for multiplication: Furthermore, any expression which evaluates to
absent is not stored in the left-hand side of an assignment statement :
POKI_RUN_COMMAND{{echo 'x=2,y=3' | mlr put '$a=$u+$v; $b=$u+$y; $c=$x+$y'}}HERE
POKI_RUN_COMMAND{{echo 'x=2,y=3' | mlr put '$a=min($x,$v);$b=max($u,$y);$c=min($u,$v)'}}HERE
Likewise, for assignment to maps, absent-valued keys or values result
in a skipped assignment.
The reasoning is as follows:
Empty values are explicit in the data so they should explicitly affect accumulations:
mlr put '@sum += $x'
should accumulate numeric x values into the sum but an empty
x, when encountered in the input data stream, should make the sum
non-numeric. To work around this you can use the
is_not_null function as follows:
mlr put 'is_not_null($x) { @sum += $x }'
Absent stream-record values should not break accumulations, since Miller
by design handles heterogenous data: the running @sum in
mlr put '@sum += $x'
should not be invalidated for records which have no x.
Absent out-of-stream-variable values are precisely what allow you to write
mlr put '@sum += $x'. Otherwise you would have to write
mlr put 'begin{@sum = 0}; @sum += $x' —
which is tolerable — but for
mlr put 'begin{...}; @sum[$a][$b] += $x'
you’d have to pre-initialize @sum for all values of $a and $b in your
input data stream, which is intolerable.
The penalty for the absent feature is that misspelled variables can be hard to find:
e.g. in mlr put 'begin{@sumx = 10}; ...; update @sumx somehow per-record; ...; end {@something = @sum * 2}'
the accumulator is spelt @sumx in the begin-block but @sum in the end-block, where since it
is absent, @sum*2 evaluates to 2. See also the section on
errors and transparency.
Since absent plus absent is absent (and likewise for other operators),
accumulations such as @sum += $x work correctly on heterogenous data,
as do within-record formulas if both operands are absent. If one operand is
present, you may get behavior you don’t desire. To work around this
— namely, to set an output field only for records which have all the
inputs present — you can use a pattern-action block with
is_present:
POKI_RUN_COMMAND{{mlr cat data/het.dkvp}}HERE
POKI_RUN_COMMAND{{mlr put 'is_present($loadsec) { $loadmillis = $loadsec * 1000 }' data/het.dkvp}}HERE
POKI_RUN_COMMAND{{mlr put '$loadmillis = (is_present($loadsec) ? $loadsec : 0.0) * 1000' data/het.dkvp}}HERE
If you’re interested in a formal description of how empty and absent
fields participate in arithmetic, here’s a table for plus (other
arithmetic/boolean/bitwise operators are similar):
POKI_RUN_COMMAND{{mlr --print-type-arithmetic-info}}HERE
String literals
You can use the following backslash escapes for strings such as between the double quotes in contexts such as
mlr filter '$name =~ "..."',
mlr put '$name = $othername . "..."',
mlr put '$name = sub($name, "...", "..."), etc.:
\a: ASCII code 0x07 (alarm/bell)
\b: ASCII code 0x08 (backspace)
\f: ASCII code 0x0c (formfeed)
\n: ASCII code 0x0a (LF/linefeed/newline)
\r: ASCII code 0x0d (CR/carriage return)
\t: ASCII code 0x09 (tab)
\v: ASCII code 0x0b (vertical tab)
\\: backslash
\": double quote
\123: Octal 123, etc. for \000 up to \377\x7f: Hexadecimal 7f, etc. for \x00 up to \xff
See also https://en.wikipedia.org/wiki/Escape_sequences_in_C.
These replacements apply only to strings you key in for the DSL expressions for filter and put:
that is, if you type \t in a string literal for a filter/put expression, it will be turned into a tab character. If you want a backslash followed by a t, then please type \\t.
However, these replacements are not done automatically within your data stream. If you wish to make these
replacements, you can do, for example, for a field named field, mlr put '$field = gsub($field, "\\t",
"\t")'. If you need to make such a replacement for all fields in your data, you should probably simply use the
system sed command.
Regular expressions
Miller lets you use regular expressions (of type POSIX.2) in the following contexts:
In mlr filter with =~ or !=~, e.g. mlr
filter '$url =~ "http.*com"'
In mlr put with sub or gsub, e.g. mlr put
'$url = sub($url, "http.*com", "")'
In mlr having-fields, e.g. mlr having-fields
--any-matching '^sda[0-9]'
In mlr cut, e.g. mlr cut -r -f '^status$,^sda[0-9]'
In mlr rename, e.g. mlr rename -r '^(sda[0-9]).*$,dev/\1'
In mlr grep, e.g. mlr --csv grep 00188555487 myfiles*.csv
Points demonstrated by the above examples:
There are no implicit start-of-string or end-of-string anchors; please
use ^ and/or $ explicitly.
Miller regexes are wrapped with double quotes rather than slashes.
The i after the ending double quote indicates a case-insensitive
regex.
Capture groups are wrapped with (...) rather than
\(...\); use \( and \) to match against parentheses.
For filter and put, if the regular expression is a string
literal (the normal case), it is precompiled at process start and reused
thereafter, which is efficient. If the regular expression is a more complex
expression, including string concatenation using ., or a column name
(in which case you can take regular expressions from input data!), then regexes
are compiled on each record which works but is less efficient. As well, in this
case there is no way to specify case-insensitive matching.
Example:
POKI_RUN_COMMAND{{cat data/regex-in-data.dat}}HERE
POKI_RUN_COMMAND{{mlr filter '$name =~ $regex' data/regex-in-data.dat}}HERE
Regex captures
Regex captures of the form \0 through \9 are supported as
follows:
Captures have in-function context for sub and gsub.
For example, the first \1,\2 pair belong to the first sub and
the second \1,\2 pair belong to the second sub:
Captures endure for the entirety of a put for the =~
and !=~ operators. For example, here the \1,\2 are set by the
=~ operator and are used by both subsequent assignment statements:
The captures are not retained across multiple puts. For example, here the
\1,\2 won’t be expanded from the regex capture:
mlr put '$a =~ "(..)_(....)' then {... something else ...} then put '$b = "left_\1"; $c = "right_\2"'
Captures are ignored in filter for the =~ and
!=~ operators. For example, there is no mechanism provided to refer to
the first (..) as \1 or to the second (....) as
\2 in the following filter statement:
mlr filter '$a =~ "(..)_(....)'
Up to nine matches are supported: \1 through \9, while
\0 is the entire match string; \15 is treated as \1
followed by an unrelated 5.
Arithmetic
Input scanning
Numbers in Miller are double-precision float or 64-bit signed integers.
Anything scannable as int, e.g 123 or 0xabcd, is treated as
an integer; otherwise, input scannable as float (4.56 or 8e9)
is treated as float; everything else is a string.
If you want all numbers to be treated as floats, then you may use
float() in your filter/put expressions (e.g. replacing $c = $a *
$b with $c = float($a) * float($b)) — or, more simply, use
mlr filter -F and mlr put -F which forces all numeric input,
whether from expression literals or field values, to float. Likewise mlr
stats1 -F and mlr step -F force integerable accumulators (such as
count) to be done in floating-point.
Conversion by math routines
For most math functions, integers are cast to float on input, and produce
float output: e.g. exp(0) = 1.0 rather than 1. The
following, however, produce integer output if their inputs are integers:
+-*///%absceilfloormaxminroundroundmsgn. As well, stats1 -a min, stats1 -a
max, stats1 -a sum, step -a delta, and step -a
rsum produce integer output if their inputs are integers.
Conversion by arithmetic operators
The sum, difference, and product of integers is again integer, except for
when that would overflow a 64-bit integer at which point Miller converts the
result to float.
The short of it is that Miller does this transparently for you so you
needn’t think about it.
Implementation details of this, for the interested: integer adds and
subtracts overflow by at most one bit so it suffices to check sign-changes.
Thus, Miller allows you to add and subtract arbitrary 64-bit signed integers,
converting only to float precisely when the result is less than -263
or greater than 263-1. Multiplies, on the other hand, can overflow
by a word size and a sign-change technique does not suffice to detect overflow.
Instead Miller tests whether the floating-point product exceeds the
representable integer range. Now, 64-bit integers have 64-bit precision while
IEEE-doubles have only 52-bit mantissas — so, there are 53 bits including
implicit leading one. The following experiment explicitly demonstrates the
resolution at this range:
64-bit integer 64-bit integer Casted to double Back to 64-bit
in hex in decimal integer
0x7ffffffffffff9ff 9223372036854774271 9223372036854773760.000000 0x7ffffffffffff800
0x7ffffffffffffa00 9223372036854774272 9223372036854773760.000000 0x7ffffffffffff800
0x7ffffffffffffbff 9223372036854774783 9223372036854774784.000000 0x7ffffffffffffc00
0x7ffffffffffffc00 9223372036854774784 9223372036854774784.000000 0x7ffffffffffffc00
0x7ffffffffffffdff 9223372036854775295 9223372036854774784.000000 0x7ffffffffffffc00
0x7ffffffffffffe00 9223372036854775296 9223372036854775808.000000 0x8000000000000000
0x7ffffffffffffffe 9223372036854775806 9223372036854775808.000000 0x8000000000000000
0x7fffffffffffffff 9223372036854775807 9223372036854775808.000000 0x8000000000000000
That is, one cannot check an integer product to see if it is precisely
greater than 263-1 or less than -263 using either integer
arithmetic (it may have already overflowed) or using double-precision (due to
granularity). Instead Miller checks for overflow in 64-bit integer
multiplication by seeing whether the absolute value of the double-precision
product exceeds the largest representable IEEE double less than 263,
which we see from the listing above is 9223372036854774784. (An alternative
would be to do all integer multiplies using handcrafted multi-word 128-bit
arithmetic. This approach is not taken.)
Quotient of integers is floating-point: 7/2 is 3.5.
Integer division is done with //: 7//2 is 3.
This rounds toward the negative.
Remainders are non-negative.