miller/doc/content-for-originality.html
John Kerl 64c63f28ad todo
2016-05-06 13:12:38 -07:00

90 lines
5.5 KiB
HTML

<p/> It isn&rsquo;t. Miller is one of many, many participants in the
online-analytical-processing culture. Other key participants include
<tt>awk</tt>, SQL, spreadsheets, etc. etc. etc. Far from being an original
concept, Miller explicitly strives to imitate several existing tools:
<p/>
<boldmaroon>Unix toolkit</boldmaroon>: Intentional similarities as described in
POKI_PUT_LINK_FOR_PAGE(feature-comparison.html)HERE.
<p/>Recipes abound for command-line data analysis using the Unix toolkit. Here are just a couple of my favorites:
<ul>
<li/> <a href="http://en.wikibooks.org/wiki/Ad_Hoc_Data_Analysis_From_The_Unix_Command_Line">http://en.wikibooks.org/wiki/Ad_Hoc_Data_Analysis_From_The_Unix_Command_Line</a>
<li/> <a href="http://www.gregreda.com/2013/07/15/unix-commands-for-data-science">http://www.gregreda.com/2013/07/15/unix-commands-for-data-science</a>
</ul>
<p/> <boldmaroon>RecordStream</boldmaroon>: Miller owes particular inspiration
to <a href="https://github.com/benbernard/RecordStream">RecordStream</a>. The
key difference is that RecordStream is a Perl-based tool for manipulating JSON
(including requiring it to separately manipulate other formats such as CSV into
and out of JSON), while Miller is fast C which handles its formats natively.
The similarities include the <tt>sort</tt>, <tt>stats1</tt> (analog of
RecordStream&rsquo;s <tt>collate</tt>), and <tt>delta</tt> operations, as well
as <tt>filter</tt> and <tt>put</tt>, and pretty-print formatting.
<p/> <boldmaroon>stats_m</boldmaroon>: A third source of lineage is my Python
<a href="https://github.com/johnkerl/scripts-math/tree/master/stats">stats_m</a>
module. This includes simple single-pass algorithms which form Miller&rsquo;s
<tt>stats1</tt> and <tt>stats2</tt> subcommands.
<p/> <boldmaroon>SQL</boldmaroon>: Fourthly, Miller&rsquo;s <tt>group-by</tt> command
name is from SQL, as is the term <tt>aggregate</tt>.
<p/> <boldmaroon>Added value</boldmaroon>:
Miller&rsquo;s added values include:
<ul>
<li> Name-indexing, compared to the Unix toolkit&rsquo;s positional indexing.
<li> Raw speed, compared to <tt>awk</tt>, RecordStream, <tt>stats_m</tt>, or various other kinds of Python/Ruby/etc. scripts one can easily create.
<li> Compact keystroking for many common tasks, with a decent amount of flexibility.
<li> Ability to handle text files on the Unix pipe, without need for creating database tables, compared to SQL databases.
<li> Various file formats, and on-the-fly format conversion.
</ul>
<p/><boldmaroon>jq</boldmaroon>: Miller does for name-indexed text what
<a href="http://stedolan.github.io/jq/">jq</a> does for JSON. If you&rsquo;re
not already familiar with <tt>jq</tt>, please <a href="http://stedolan.github.io/jq/">check it out!</a>.
<p/><boldmaroon>What about DOTADIW?</boldmaroon> One of the key points of the
<a href="http://en.wikipedia.org/wiki/Unix_philosophy">Unix philosophy</a> is
that a tool should do one thing and do it well. Hence <tt>sort</tt> and
<tt>cut</tt> do just one thing. Why does Miller put <tt>awk</tt>-like
processing, a few SQL-like operations, and statistical reduction all into one
tool (see also POKI_PUT_LINK_FOR_PAGE(reference.html)HERE)? This is a fair
question. First note that many standard tools, such as <tt>awk</tt> and
<tt>perl</tt>, do quite a few things &mdash; as does <tt>jq</tt>. But I could
have pushed for putting format awareness and name-indexing options into
<tt>cut</tt>, <tt>awk</tt>, and so on (so you could do <tt>cut -f
hostname,uptime</tt> or <tt>awk '{sum += $x*$y}END{print sum}'</tt>). Patching
<tt>cut</tt>, <tt>sort</tt>, etc. on multiple operating systems is a
non-starter in terms of uptake. Moreover, it makes sense for me to have Miller
be a tool which collects together format-aware record-stream processing into
one place, with good reuse of Miller-internal library code for its various
features.
<p/><boldmaroon>Why not use Perl/Python/Ruby etc.?</boldmaroon> Maybe you
should. With those tools you&rsquo;ll get far more expressive power, and
sufficiently quick turnaround time for small-to-medium-sized data. Using
Miller you&rsquo;ll get something less than a complete programming language,
but which is fast, with moderate amounts of flexibility and much less
keystroking.
<p/>When I was first developing Miller I made a survey of several languages.
Using low-level implementation languages like C, Go, Rust, and Nim, I&rsquo;d
need to create my own domain-specific language (DSL) which would always be less
featured than a full programming language, but I&rsquo;d get better
performance. Using high-level interpreted languages such as Perl/Python/Ruby
I&rsquo;d get the language&rsquo;s <tt>eval</tt> for free and I wouldn&rsquo;t
need a DSL; Miller would have mainly been a set of format-specific I/O hooks.
If I&rsquo;d gotten good enough performance from the latter I&rsquo;d have done
it without question and Miller would be far more flexible. But C won the
performance criteria by a landslide so we have Miller in C with a custom DSL.
<p/> <boldmaroon>No, really, why one more command-line data-manipulation
tool?</boldmaroon> I wrote Miller because I was frustrated with tools like
<tt>grep</tt>, <tt>sed</tt>, and so on being <i>line-aware</i> without being
<i>format-aware</i>. The single most poignant example I can think of is seeing
people grep data lines out of their CSV files and sadly losing their header
lines. While some lighter-than-SQL processing is very nice to have, at core I
wanted the format-awareness of <a
href="https://github.com/benbernard/RecordStream">RecordStream</a> combined
with the raw speed of the Unix toolkit. Miller does precisely that.