https://github.com/dkogan/vnlog

Process labelled tabular ASCII data using normal UNIX tools
https://github.com/dkogan/vnlog

Last synced: about 1 year ago
JSON representation

Process labelled tabular ASCII data using normal UNIX tools

Host: GitHub
URL: https://github.com/dkogan/vnlog
Owner: dkogan
Created: 2017-12-23T08:25:47.000Z (over 8 years ago)
Default Branch: master
Last Pushed: 2025-02-15T03:49:14.000Z (over 1 year ago)
Last Synced: 2025-03-28T23:08:40.111Z (about 1 year ago)
Language: Perl
Homepage:
Size: 836 KB
Stars: 161
Watchers: 7
Forks: 6
Open Issues: 0
Metadata Files:
- Readme: README.org
- Changelog: Changes

Awesome Lists containing this project

README

* Talk

I just gave a talk about this at [[https://www.socallinuxexpo.org/scale/17x][SCaLE 17x]]. Here are the [[https://www.youtube.com/watch?v=Qvb_uNkFGNQ&t=12830s][video of the talk]] and
the [[https://github.com/dkogan/talk-feedgnuplot-vnlog/blob/master/feedgnuplot-vnlog.org]["slides"]].

* Summary

Vnlog ("vanilla-log") is a toolkit for manipulating tabular ASCII data with
labelled fields using normal UNIX tools. If you regularly use =awk= and =sort=
and =uniq= and others, these tools will make you infinitely more powerful. The
vnlog tools /extend/, rather than replace the standard tooling, so minimal
effort is required to learn and use these tools.

Everything assumes a trivially simple log format:

- A whitespace-separated table of ASCII human-readable text
- A =#= character starts a comment that runs to the end of the line (like in
many scripting languages)
- The first line that begins with a single =#= (not =##= or =#!=) is a /legend/,
naming each column. This is required, and the field names that appear here are
referenced by all the tools.
- Empty fields reported as =-=

This describes 99% of the format, with some extra details [[#format-details][below]]. Example:

#+BEGIN_EXAMPLE
#!/usr/bin/whatever
# a b c
1 2 3
## comment
4 5 6
#+END_EXAMPLE

Such data can be processed directly with almost any existing tool, and /this/
toolkit allows the user to manipulate this data in a nicer way by relying on
standard UNIX tools. The core philosophy is to avoid creating new knowledge as
much as possible. Consequently, the vnlog toolkit relies /heavily/ on existing
(and familiar!) tools and workflows. As such, the toolkit is small, light, and
has a /very/ friendly learning curve.

* Synopsis

I have [[https://raw.githubusercontent.com/dkogan/vnlog/master/dji-tsla.tar.gz][two sets of historical stock data]], from the start of 2018 until now
(2018/11):

#+BEGIN_SRC sh :results output :exports both
< dji.vnl head -n 4
#+END_SRC

#+RESULTS:
: # Date Open High Low Close AdjClose Volume
: 2018-11-15 25061.48 25354.56 24787.79 25289.27 25289.27 383292840
: 2018-11-14 25388.08 25501.29 24935.82 25080.50 25080.50 384240000
: 2018-11-13 25321.21 25511.03 25193.78 25286.49 25286.49 339690000

And

#+BEGIN_SRC sh :results output :exports both
< tsla.vnl head -n 4
#+END_SRC

#+RESULTS:
: # Date Open High Low Close AdjClose Volume
: 2018-11-15 342.33 348.58 339.04 348.44 348.44 4486339
: 2018-11-14 342.70 347.11 337.15 344.00 344.00 5036300
: 2018-11-13 333.16 344.70 332.20 338.73 338.73 5448600

I can add whitespace to make the headers more legible by humans:

#+BEGIN_SRC sh :results output :exports both
< dji.vnl head -n 4 | vnl-align
#+END_SRC

I can pull out the closing prices:

#+BEGIN_SRC sh :results output :exports both
< dji.vnl vnl-filter -p Close | head -n4
#+END_SRC

#+RESULTS:
: # Close
: 25289.27
: 25080.50
: 25286.49

=vnl-filter= is primarily a wrapper around =awk= or =perl=, allowing the user to
reference columns by name. I can then plot the closing prices:

#+BEGIN_SRC sh :results file link :exports both
< dji.vnl vnl-filter -p Close |
feedgnuplot --lines --unset grid
#+END_SRC

#+RESULTS:
[[file:guide-1.svg]]

Here I kept /only/ the closing price column, so the x-axis is just the row
index. The data was in reverse chronological order, so this plot is also in
reverse chronological order. Let's fix that:

#+BEGIN_SRC sh :results file link :exports both
< dji.vnl vnl-sort -k Date |
vnl-filter -p Close |
feedgnuplot --lines --unset grid
#+END_SRC

#+RESULTS:
[[file:guide-2.svg]]

The =vnl-sort= tool (and most of the other =vnl-xxx= tools) are wrappers around
the core tools already available on the system (such as =sort=, in this case).
With the primary difference being reading/writing vnlog, and referring to
columns by name.

We now have the data in the correct order, but it'd be nice to see the actual
dates on the x-axis. While we're at it, let's label the axes too:

#+BEGIN_SRC sh :results output :exports both
< dji.vnl vnl-filter -p Date,Close | head -n4
#+END_SRC

#+RESULTS:
: # Date Close
: 2018-11-15 25289.27
: 2018-11-14 25080.50
: 2018-11-13 25286.49

#+BEGIN_SRC sh :results file link :exports both
< dji.vnl vnl-sort -k Date |
vnl-filter -p Date,Close |
feedgnuplot --lines --unset grid --timefmt %Y-%m-%d --domain \
--xlabel 'Date' --ylabel 'Price ($)'
#+END_SRC

#+RESULTS:
[[file:guide-3.svg]]

What was the highest value of the Dow-Jones index, and when did it happen?

#+BEGIN_SRC sh :results output :exports both
< dji.vnl vnl-sort -rgk Close |
head -n2 |
vnl-align
#+END_SRC

#+RESULTS:
: # Date Open High Low Close AdjClose Volume
: 2018-10-03 26833.47 26951.81 26789.08 26828.39 26828.39 280130000

Alrighty. Looks like the high was in October. Let's zoom in on that month:

#+BEGIN_SRC sh :results file link :exports both
< dji.vnl vnl-sort -k Date |
vnl-filter 'Date ~ /2018-10/' -p Date,Close |
feedgnuplot --lines --unset grid --timefmt %Y-%m-%d --domain \
--xlabel 'Date' --ylabel 'Price ($)'
#+END_SRC

#+RESULTS:
[[file:guide-4.svg]]

OK. Is this thing volatile? What was the largest single-day gain?

#+BEGIN_SRC sh :results output :exports both
< dji.vnl vnl-filter -p '.,d=diff(Close)' |
head -n4 |
vnl-align
#+END_SRC

#+RESULTS:
: # Date Open High Low Close AdjClose Volume d
: 2018-11-15 25061.48 25354.56 24787.79 25289.27 25289.27 383292840 -
: 2018-11-14 25388.08 25501.29 24935.82 25080.50 25080.50 384240000 -208.77
: 2018-11-13 25321.21 25511.03 25193.78 25286.49 25286.49 339690000 205.99

#+BEGIN_SRC sh :results output :exports both
< dji.vnl vnl-filter -p '.,d=diff(Close)' |
vnl-sort -rgk d |
head -n2 |
vnl-align
#+END_SRC

#+RESULTS:
: # Date Open High Low Close AdjClose Volume d
: 2018-02-02 26061.79 26061.79 25490.66 25520.96 25520.96 522880000 1175.21

Whoa. So the best single-gain day was 2018-02-02: the dow gained 1175.21 points
between closing on Feb 1 and Feb 2. But it actually lost ground that day! What
if I looked at the difference between the opening and closing in a single day?

#+BEGIN_SRC sh :results output :exports both
< dji.vnl vnl-filter -p '.,d=Close-Open' |
vnl-sort -rgk d |
head -n2 |
vnl-align
#+END_SRC

#+RESULTS:
: # Date Open High Low Close AdjClose Volume d
: 2018-02-06 24085.17 24946.23 23778.74 24912.77 24912.77 823940000 827.6

I guess by that metric 2018-02-06 was better. Let's join the Dow-jones index
data and the TSLA data, and let's look at them together:

#+BEGIN_SRC sh :results output :exports both
vnl-join --vnl-autosuffix dji.vnl tsla.vnl -j Date |
head -n4 |
vnl-align
#+END_SRC

#+RESULTS:
: # Date Open_dji High_dji Low_dji Close_dji AdjClose_dji Volume_dji Open_tsla High_tsla Low_tsla Close_tsla AdjClose_tsla Volume_tsla
: 2018-11-15 25061.48 25354.56 24787.79 25289.27 25289.27 383292840 342.33 348.58 339.04 348.44 348.44 4486339
: 2018-11-14 25388.08 25501.29 24935.82 25080.50 25080.50 384240000 342.70 347.11 337.15 344.00 344.00 5036300
: 2018-11-13 25321.21 25511.03 25193.78 25286.49 25286.49 339690000 333.16 344.70 332.20 338.73 338.73 5448600

#+BEGIN_SRC sh :results output :exports both
vnl-join --vnl-autosuffix dji.vnl tsla.vnl -j Date |
vnl-filter -p '^Close' |
head -n4 |
vnl-align
#+END_SRC

#+RESULTS:
: # Close_dji Close_tsla
: 25289.27 348.44
: 25080.50 344.00
: 25286.49 338.73

#+BEGIN_SRC sh :results file link :exports both
vnl-join --vnl-autosuffix dji.vnl tsla.vnl -j Date |
vnl-filter -p '^Close' |
feedgnuplot --domain --points --unset grid \
--xlabel 'DJI price ($)' --ylabel 'TSLA price ($)'
#+END_SRC

#+RESULTS:
[[file:guide-5.svg]]

Huh. Apparently there's no obvious, strong correlation between TSLA and
Dow-Jones closing prices. And we saw that with just a few shell commands,
without dropping down into a dedicated analysis system.

* Build and installation
vnlog is a part of Debian/buster and Ubuntu/cosmic (18.10) and later. On those
boxes you can simply

#+BEGIN_EXAMPLE
$ sudo apt install vnlog libvnlog-dev libvnlog-perl python3-vnlog
#+END_EXAMPLE

to get the binary tools, the C API, the perl and python3 interfaces
respectively.

** Install on non-Debian boxes
Most of this is written in an interpreted language, so there's nothing to build
or install, and you can run the tools directly from the source tree:

#+BEGIN_EXAMPLE
$ git clone https://github.com/dkogan/vnlog.git
$ cd vnlog
$ ./vnl-filter .....
#+END_EXAMPLE

The python and perl libraries can be run from the tree by setting the
=PYTHONPATH= and =PERL5LIB= environment variables respectively. For the C
library, you should =make=, and then point your =CFLAGS= and =LDLIBS= and
=LD_LIBRARY_PATH= to the local tree.

If you do want to install to some arbitrary location to simplify the paths, do
this:

#+BEGIN_EXAMPLE
$ make
$ PREFIX=/usr/local make install
#+END_EXAMPLE

This will install /all/ the components into =/usr/local=.

* Description
Vnlog data is nicely readable by both humans and machines. Any time your
application invokes =printf()= for either diagnostics or logging, consider
writing out vnlog-formatted data. You retain human readability, but gain the
power all the =vnl-...= tools provide.

Vnlog tools are designed to be very simple and light. There's an ever-growing
list of other tools that do vaguely the same thing. Some of these:

- https://github.com/BurntSushi/xsv
- https://csvkit.readthedocs.io/
- https://github.com/johnkerl/miller
- https://github.com/jqnatividad/qsv
- https://github.com/greymd/teip
- https://github.com/eBay/tsv-utils-dlang
- https://www.gnu.org/software/datamash/
- https://stedolan.github.io/jq/
- https://github.com/benbernard/RecordStream
- https://github.com/dinedal/textql
- https://www.visidata.org/
- http://harelba.github.io/q/
- https://github.com/BatchLabs/charlatan
- https://github.com/dbohdan/sqawk

Many of these provide facilities to run various analyses, and others focus on
data types that aren't just a table (json for instance). Vnlog by contrast
doesn't analyze anything, and targets the most trivial possible data format.
This makes it very easy to run any analysis you like in any tool you like. The
main envisioned use case is one-liners, and the tools are geared for that
purpose. The above mentioned tools are much more powerful than vnlog, so they
could be a better fit for some use cases. I claim that

- 90% of the time you want to do simple things, and vnlog is a great fit for the
task
- If you really do need to do something complex, you shouldn't be in the shell
writing oneliners anymore, and a fully-fledged analysis system (numpy, etc) is
more appropriate

In the spirit of doing as little as possible, the provided tools are wrappers
around tools you already have and are familiar with. The provided tools are:

- =vnl-filter= is a tool to select a subset of the rows/columns in a vnlog
and/or to manipulate the contents. This is an =awk= wrapper where the fields
can be referenced by name instead of index. 20-second tutorial:

#+BEGIN_SRC sh :results none :exports code
vnl-filter -p col1,col2,colx=col3+col4 'col5 > 10' --has col6
#+END_SRC

will read the input, and produce a vnlog with 3 columns: =col1= and =col2=
from the input, and a column =colx= that's the sum of =col3= and =col4= in the
input. Only those rows for which /both/ =col5 > 10= is true /and/ that have a
non-null value for =col6= will be output. A null entry is signified by a
single =-= character.

#+BEGIN_SRC sh :results none :exports code
vnl-filter --eval '{s += x} END {print s}'
#+END_SRC

#+RESULTS:

will evaluate the given awk program on the input, but the column names work as
you would hope they do: if the input has a column named =x=, this would
produce the sum of all values in this column.

- =vnl-sort=, =vnl-uniq=, =vnl-join=, =vnl-tail=, =vnl-ts= are wrappers around
the corresponding commandline tools. These work exactly as you would expect
also: the columns can be referenced by name, and the legend comment is handled
properly. These are wrappers, so all the commandline options those tools have
"just work" (except options that don't make sense in the context of vnlog). As
an example, =vnl-tail -f= will follow a log: data will be read by =vnl-tail=
as it is written into the log (just like =tail -f=, but handling the legend
properly). And you already know how to use these tools without even reading
the manpages! Note: I use the Linux kernel and the tools from GNU Coreutils
exclusively, but this all has been successfully tested on FreeBSD and OSX
also. Please let me know if something doesn't work.

- =vnl-align= aligns vnlog columns for easy interpretation by humans. The
meaning is unaffected

- =Vnlog::Parser= is a simple perl library to read a vnlog

- =vnlog= is a simple python library to read a vnlog. Both python2 and python3
are supported

- =libvnlog= is a C library to simplify reading and writing a vnlog. Clearly all
you /really/ need for writing is =printf()=, but this is useful if we have
lots of columns, many containing null values in any given row, and/or if we
have parallel threads writing to a log. In my usage I have hundreds of columns
of sparse data, so this is handy

- =vnl-make-matrix= converts a one-point-per-line vnlog to a matrix of data.
I.e.

#+BEGIN_EXAMPLE
$ cat dat.vnl
# i j x
0 0 1
0 1 2
0 2 3
1 0 4
1 1 5
1 2 6
2 0 7
2 1 8
2 2 9
3 0 10
3 1 11
3 2 12

$ < dat.vnl vnl-filter -p i,x | vnl-make-matrix --outdir /tmp
Writing to '/tmp/x.matrix'

$ cat /tmp/x.matrix
1 2 3
4 5 6
7 8 9
10 11 12
#+END_EXAMPLE

All the tools have manpages that contain more detail. And more tools will
probably be added with time.

* Format details
The high-level description of the vnlog format from [[#Summary][above]] is sufficient to
read/write "normal" vnlog data, but there are a few corner cases that should be
mentioned. To reiterate, the format description from above describes vnlog as:

For a few years now I've been using these tools myself, and supporting others as
they were passing vnlog data around. In the process I've encountered some
slightly-weird data, and patched the tools to accept it. So today the included
vnlog tools are /very/ permissive, and accept any vnlog data that can possibly
be accepted. Other vnlog tools may not be quite as permissive, and may not be
able to interpret "weird" data. Points of note, describing the included vnlog
tools:

- Leading and trailing whitespace is ignored. Everywhere. So this data file will
be read properly, with the =x= column containing 1 and 3:

#+begin_example
# x y
1 2
3 4
#+end_example

- Empty (or whitespace-only) lines anywhere are ignored, and treated as a
comment
- An initial =#= comment without field names is treated as a comment, and we
continue looking for the legend in the following lines. So this data file will
be read properly:

#+begin_example
## comment
#
# x y
1 2
3 4
#+end_example

- Trailing comments are supported, like in most scripting languages. So this
data file will be read properly:

#+begin_example
# x y
1 2 # comment
3 4
#+end_example

- Field names are /very/ permissive: anything that isn't whitespace is
supported. So this data file will be read properly:

#+begin_example
# x y # 1+ -
1 2 3 4 5
11 12 13 14 15
#+end_example

We can pull out the =#= and =1+= and =-= columns:

#+begin_src sh
vnl-filter -p '#,1+,-'
#+end_src

And we can even operate on them, if we use whitespace to indicate field
boundaries:

#+begin_src sh
vnl-filter -p 'x=1+ + 5'
#+end_src

Note that this implies that trailing comments in a legend line are /not/
supported: the extra =#= characters will be used for field names. Field names
containing =,= or === are currently not accepted by =vnl-filter=, but /are/
accepted by the other tools (=vnl-sort= and such). I'll make =vnl-filter= able
to work with those field names too, eventually, but as a user, the simplest
thing to do is to not pass around data with such field names.

- Duplicated labels are supported whenever possible. So

#+begin_example
# x y z z
1 2 3 4
11 12 13 14
#+end_example

will work just fine, unless we're operating on =z=. With this data, both of
these commands work:

#+begin_src sh
vnl-filter -p x
vnl-filter -p z
#+end_src

Picking =z= selects both of the =z= columns. But neither of these commands can
work with the non-unique =z= column:

#+begin_src sh
vnl-filter -p s=z+1
vnl-sort -k z
#+end_src

* Workflows and recipes
** Storing disjoint data

A common use case is a complex application that produces several semi-related
subsets of data at once. Example: a moving vehicle is reporting both its own
position and the observed positions of other vehicles; at any given time any
number of other vehicles may be observed. Two equivalent workflows are possible:

- a single unified vnlog stream for /all/ the data
- several discrete vnlog streams for each data subset

Both are valid approaches

*** One unified vnlog stream
Here the application produces a /single/ vnlog that contains /all/ the columns,
from /all/ the data subsets. In any given row, many of the columns will be empty
(i.e. contain only =-= ). For instance, a row describing a vehicle own position
will not have data about any observations, and vice versa. It is inefficient to
store all the extra =-= but it makes many things much nicer, so it's often worth
it. =vnl-filter= can be used to pull out the different subsets. Sample
=joint.vnl=:

#+BEGIN_EXAMPLE
# time x_self x_observation
1 10 -
2 20 -
2 - 100
3 30 -
3 - 200
3 - 300
#+END_EXAMPLE

Here we have 3 instances in time. We have no observations at =time= 1, one
observation at =time= 2, and two observations at =time= 3. We can use
=vnl-filter= to pull out the data we want:

#+BEGIN_EXAMPLE
$ < joint.vnl vnl-filter -p time,self

# time x_self
1 10
2 20
2 -
3 30
3 -
3 -
#+END_EXAMPLE

If we only care about our own positions, the =+= modifier in picked columns in
=vnl-filter= is very useful here:

#+BEGIN_EXAMPLE
$ < joint.vnl vnl-filter -p time,+self

# time x_self
1 10
2 20
3 30

$ < joint.vnl vnl-filter -p time,+observation

# time x_observation
2 100
3 200
3 300
#+END_EXAMPLE

Note that the default is =--skipempty=, so if we're /only/ looking at =x_self=
for instance, then we don't even need to =+= modifier:

#+begin_example
$ < joint.vnl vnl-filter -p self

# x_self
10
20
30
#+end_example

Also, note that the =vnlog= C interface works very nicely to produce these
datafiles:

- You can define lots and lots of columns, but only fill some of them before
calling =vnlog_emit_record()=. The rest will be set to =-=.
- You can create multiple contexts for each type of data, and you can populate
them with data independently. And when calling =vnlog_emit_record_ctx()=,
you'll get a record with data for just that context.

*** Several discrete vnlog streams

Conversely, the application can produce /separate/ vnlog streams for /each/
subset of data. Depending on what is desired, exactly, =vnl-join= can be used to
re-join them:

#+BEGIN_EXAMPLE
$ cat self.vnl

# time x_self
1 10
2 20
3 30

$ cat observations.vnl

# time x_observation
2 100
3 200
3 300

$ vnl-join -j time -a- self.vnl observations.vnl

# time x_self x_observation
1 10 -
2 20 100
3 30 200
3 30 300
#+END_EXAMPLE

** Data statistics

A common need is to compute basic statistics from your data. Many of the
alternative toolkits listed above provide built-in facilities to do this, but
vnlog does not: it's meant to be unixy, where each tool has very limited scope.
Thus you can either do this with =awk= like you would normally, or you can use
other standalone tools to perform the needed computations. For instance, I can
generate some data:

#+BEGIN_EXAMPLE
$ seq 2 100 | awk 'BEGIN {print "# x"} {print log($1)}' > /tmp/log.vnl
#+END_EXAMPLE

Then I can compute the mean with =awk=:

#+BEGIN_EXAMPLE
$ < /tmp/log.vnl vnl-filter --eval '{sum += x} END {print sum/NR}'
3.67414
#+END_EXAMPLE

Or I can compute the mean (and other stuff) with a separate standalone tool:

#+BEGIN_EXAMPLE
$ < /tmp/log.vnl ministat
x
+----------------------------------------------------------------------------+
| xx |
| x xxxxxxx |
| xx xxxxxxxxxxxx|
| x x xxxxxxxxxxxxxxxxxxxxxxx|
|x x x x x x x x x xx xx xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx|
| |_______________A____M___________| |
+----------------------------------------------------------------------------+
N Min Max Median Avg Stddev
x 99 0.693147 4.60517 3.93183 3.6741353 0.85656382
#+END_EXAMPLE

=ministat= is not a part of the vnlog toolkit, but the vnlog format is generic
so it works just fine.

** Powershell-style filtering of common shell commands

Everything about vnlog is generic and simple, so it's easy to use it to process
data that wasn't originally meant to be used this way. For instance filtering
the output of =ls -l= to report only file names and sizes, skipping directories,
and sorting by file sizes:

#+BEGIN_EXAMPLE
$ ls -l

total 320
-rw-r--r-- 1 dima dima 5044 Aug 25 15:04 Changes
-rw-r--r-- 1 dima dima 12749 Aug 25 15:04 Makefile
-rw-r--r-- 1 dima dima 69789 Aug 25 15:04 README.org
-rw-r--r-- 1 dima dima 33781 Aug 25 15:04 README.template.org
-rw-r--r-- 1 dima dima 5359 Aug 25 15:04 b64_cencode.c
drwxr-xr-x 4 dima dima 4096 Aug 25 15:04 completions
drwxr-xr-x 3 dima dima 4096 Aug 25 15:04 lib
drwxr-xr-x 3 dima dima 4096 Aug 25 15:04 packaging
drwxr-xr-x 2 dima dima 4096 Aug 25 15:04 test
-rwxr-xr-x 1 dima dima 5008 Aug 25 15:04 vnl-align
-rwxr-xr-x 1 dima dima 56637 Aug 25 15:04 vnl-filter
-rwxr-xr-x 1 dima dima 5678 Aug 25 15:04 vnl-gen-header
-rwxr-xr-x 1 dima dima 29815 Aug 25 15:04 vnl-join
-rwxr-xr-x 1 dima dima 3631 Aug 25 15:04 vnl-make-matrix
-rwxr-xr-x 1 dima dima 8372 Aug 25 15:04 vnl-sort
-rwxr-xr-x 1 dima dima 5822 Aug 25 15:04 vnl-tail
-rwxr-xr-x 1 dima dima 4439 Aug 25 15:04 vnl-ts
-rw-r--r-- 1 dima dima 559 Aug 25 15:04 vnlog-base64.h
-rw-r--r-- 1 dima dima 8169 Aug 25 15:04 vnlog.c
-rw-r--r-- 1 dima dima 12677 Aug 25 15:04 vnlog.h

$ (echo '# permissions num_links user group size month day time name';
ls -l | tail -n +2) |
vnl-filter 'permissions !~ "^d"' -p name,size |
vnl-sort -gk size |
vnl-align

# name size
vnlog-base64.h 559
vnl-make-matrix 3631
vnl-ts 4439
vnl-align 5008
Changes 5044
b64_cencode.c 5359
vnl-gen-header 5678
vnl-tail 5822
vnlog.c 8169
vnl-sort 8372
vnlog.h 12677
Makefile 12749
vnl-join 29815
README.template.org 33781
vnl-filter 56637
README.org 69789
#+END_EXAMPLE

With a bit of shell manipulation, these tools can be applied to a whole lot of
different data streams that know nothing of vnlog.

* C interface
** Writing vnlog files
*** Basic usage
For most uses, vnlog files are simple enough to be generated with plain prints.
But then each print statement has to know which numeric column we're populating,
which becomes effortful with many columns. In my usage it's common to have a
large parallelized C program that's writing logs with hundreds of columns where
any one record would contain only a subset of the columns. In such a case, it's
helpful to have a library that can output the log files. This is available.
Basic usage looks like this:

In a shell:

#+BEGIN_SRC sh :results none :exports code
vnl-gen-header 'int w' 'uint8_t x' 'char* y' 'double z' 'void* binary' > vnlog_fields_generated.h
#+END_SRC

#+RESULTS:

In a C program test.c:

#+BEGIN_SRC C
#include "vnlog_fields_generated.h"

int main()
{
vnlog_emit_legend();

vnlog_set_field_value__w(-10);
vnlog_set_field_value__x(40);
vnlog_set_field_value__y("asdf");
vnlog_emit_record();

vnlog_set_field_value__z(0.3);
vnlog_set_field_value__x(50);
vnlog_set_field_value__w(-20);
vnlog_set_field_value__binary("\x01\x02\x03", 3);
vnlog_emit_record();

vnlog_set_field_value__w(-30);
vnlog_set_field_value__x(10);
vnlog_set_field_value__y("whoa");
vnlog_set_field_value__z(0.5);
vnlog_emit_record();

return 0;
}
#+END_SRC

Then we build and run, and we get

#+BEGIN_EXAMPLE
$ cc -o test test.c -lvnlog

$ ./test

# w x y z binary
-10 40 asdf - -
-20 50 - 0.2999999999999999889 AQID
-30 10 whoa 0.5 -
#+END_EXAMPLE

The binary field in base64-encoded. This is a rarely-used feature, but sometimes
you really need to log binary data for later processing, and this makes it
possible.

So you

1. Generate the header to define your columns

2. Call =vnlog_emit_legend()=

3. Call =vnlog_set_field_value__...()= for each field you want to set in that
row.

4. Call =vnlog_emit_record()= to write the row and to reset all fields for the
next row. Any fields unset with a =vnlog_set_field_value__...()= call are
written as null: =-=

This is enough for 99% of the use cases. Things get a bit more complex if we
have have threading or if we have multiple vnlog ouput streams in the same
program. For both of these we use vnlog /contexts/.

*** Contexts

To support independent writing into the same vnlog (possibly by multiple
threads; this is reentrant), each log-writer should create a context, and use it
when talking to vnlog. The context functions will make sure that the fields in
each context are independent and that the output records won't clobber each
other:

#+BEGIN_SRC C
void child_writer( // the parent context also writes to this vnlog. Pass NULL to
// use the global one
struct vnlog_context_t* ctx_parent )
{
struct vnlog_context_t ctx;
vnlog_init_child_ctx(&ctx, ctx_parent);

while(records)
{
vnlog_set_field_value_ctx__xxx(&ctx, ...);
vnlog_set_field_value_ctx__yyy(&ctx, ...);
vnlog_set_field_value_ctx__zzz(&ctx, ...);
vnlog_emit_record_ctx(&ctx);
}

vnlog_free_ctx(&ctx); // required only if we have any binary fields
}
#+END_SRC

If we want to have multiple independent vnlog writers to /different/ streams
(with different columns and legends), we do this instead:

=file1.c=:
#+BEGIN_SRC C
#include "vnlog_fields_generated1.h"

void f(void)
{
// Write some data out to the default context and default output (STDOUT)
vnlog_emit_legend();
...
vnlog_set_field_value__xxx(...);
vnlog_set_field_value__yyy(...);
...
vnlog_emit_record();
}
#+END_SRC

=file2.c=:
#+BEGIN_SRC C
#include "vnlog_fields_generated2.h"

void g(void)
{
// Make a new session context, send output to a different file, write
// out legend, and send out the data
struct vnlog_context_t ctx;
vnlog_init_session_ctx(&ctx);
FILE* fp = fopen(...);
vnlog_set_output_FILE(&ctx, fp);
vnlog_emit_legend_ctx(&ctx);
...
vnlog_set_field_value__a(...);
vnlog_set_field_value__b(...);
...
vnlog_free_ctx(&ctx); // required only if we have any binary fields
vnlog_emit_record();
}
#+END_SRC

Note that it's the user's responsibility to make sure the new sessions go to a
different =FILE= by invoking =vnlog_set_output_FILE()=. Furthermore, note that
the included =vnlog_fields_....h= file defines the fields we're writing to; and
if we have multiple different vnlog field definitions in the same program (as in
this example), then the different writers /must/ live in different source files.
The compiler will barf if you try to =#include= two different
=vnlog_fields_....h= files in the same source.

*** Remaining APIs

- =vnlog_printf(...)= and =vnlog_printf_ctx(ctx, ...)= write to a pipe like
=printf()= does. This exists primarily for comments.

- =vnlog_clear_fields_ctx(ctx, do_free_binary)= clears out the data in a context
and makes it ready to be used for the next record. It is rare for the user to
have to call this manually. The most common case is handled automatically
(clearing out a context after emitting a record). One area where this is useful
is when making a copy of a context:

#+BEGIN_SRC C
struct vnlog_context_t ctx1;
// .... do stuff with ctx1 ... add data to it ...

struct vnlog_context_t ctx2 = ctx1;
// ctx1 and ctx2 now both have the same data, and the same pointers to
// binary data. I need to get rid of the pointer references in ctx1

vnlog_clear_fields_ctx(&ctx1, false);
#+END_SRC

- =vnlog_free_ctx(ctx)= frees memory for an vnlog context. Do this before
throwing the context away. Currently this is only needed for context that have
binary fields, but this should be called for all contexts anyway, in case this
changes in a later revision

** Reading vnlog files
The basic usage goes like this:

#+begin_src c
#include
#include
#include
bool parse_vnlog(const char* filename)
{
FILE* fp = fopen(filename);
if(fp == NULL)
return false;

vnlog_parser_t ctx;
if(VNL_OK != vnlog_parser_init(&ctx, fp))
return false;

// String in the "time" column for the most-recently-parsed row
const char*const* time_record = vnlog_parser_record_from_key(&ctx, "time");
if(time_record == NULL)
{
vnlog_parser_free(&ctx);
return false;
}

int i_record = 0;
vnlog_parser_result_t result;
while(VNL_OK == (result = vnlog_parser_read_record(&ctx, fp)))
{
for(int i=0; i array(['image1.png', 'image2.png'], dtype=' array([[1, 2, 5],
[3, 4, 1]])

print(arr['temperature'])
---> array([34., 35.])
#+end_example

Notes:

- The given structured dtype defines both how to organize the data, and which
data to extract. So it can be used to read in only a subset of the available
columns. Here I could have omitted the 'temperature' column, for instance

- Sub-arrays are allowed. In the example I could say either

#+begin_src python
dtype = np.dtype([ ('image', 'U16'),
('x y z', int, (3,)),
('temperature', float), ])
#+end_src

#+begin_src python
dtype = np.dtype([ ('image', 'U16'),
('x', int),
('y', int),
('z', int),
('temperature', float), ])
#+end_src

The latter would read =x=, =y=, =z= into separate, individual arrays. Sometime
we want this, sometimes not.

- Nested structured dtypes are not allowed. Fields inside other fields are not
supported, since it's not clear how to map that to a flat vnlog legend

- If a structured dtype is given, =slurp()= returns the array only, since the
field names are already available in the dtype

* numpy interface
If we need to read data into numpy specifically, nicer tools are available than
the generic =vnlog= Python module. The built-in =numpy.loadtxt= =numpy.savetxt=
functions work well (with the caveat that =numpy.loadtxt()= should be followed
by =numpysane.atleast_dims(..., -2)= to make sure that a data array of shape
=(Nrows,Ncols)= is returned even if =Nrows==1=. For example to write to standard
output a vnlog with fields =a=, =b= and =c=:

#+BEGIN_SRC python
numpy.savetxt(sys.stdout, array, fmt="%g", header="a b c")
#+END_SRC

Note that numpy automatically adds the =#= to the header. To read a vnlog from a
file on disk, do something like

#+BEGIN_SRC python
array = numpysane.atleast_dims(numpy.loadtxt('data.vnl'), -2)
#+END_SRC

These functions know that =#= lines are comments, but don't interpret anything
as field headers. That's easy to do, so I'm not providing any helper libraries.
I might do that at some point, but in the meantime, patches are welcome.

* Compatibility

I use GNU/Linux-based systems exclusively, but everything has been tested
functional on FreeBSD and OSX in addition to Debian, Ubuntu and CentOS. I can
imagine there's something I missed when testing on non-Linux systems, so please
let me know if you find any issues.

* Caveats and bugs

These tools are meant to be simple, so some things are hard requirements. A big
one is that columns are whitespace-separated. There is /no/ mechanism for
escaping or quoting whitespace into a single field. I think supporting something
like that is more trouble than it's worth.

* Manpages
** vnl-filter
#+BEGIN_EXAMPLE
NAME
vnl-filter - filters vnlogs to select particular rows, fields

SYNOPSIS
$ cat run.vnl

# time x y z temperature
3 1 2.3 4.8 30
4 1.1 2.2 4.7 31
6 1 2.0 4.0 35
7 1 1.6 3.1 42

$ = 35' | vnl-align

# time x y z temperature
6 1 2.0 4.0 35
7 1 1.6 3.1 42

$ 10'

would select only those rows whose "size" column contains a value > 10.
See the detailed description of matches expressions below for more
detail.

Context lines
"vnl-filter" supports the context output options ("-A", "-B" and "-C")
exactly like the "grep" tool. I.e to print out all rows whose "size"
column contains a value > 10 *but also* include the 3 rows immediately
before *and* after such matching rows, do this:

vnl-filter -C3 'size > 10'

"-B" reports the rows *before* matching ones and "-A" the rows *after*
matching ones. "-C" reports both. Note that this applies *only* to
*matches* expressions: records skipped because they fail "--has" or
"--skipempty" are *not* included in contextual output.

Backend choice
By default, the parsing of arguments and the legend happens in perl,
which then constructs a simple awk script, and invokes "mawk" to
actually read the data and to process it. This is done because awk is
lighter weight and runs faster, which is important because our data sets
could be quite large. We default to "mawk" specifically, since this is a
simpler implementation than "gawk", and runs much faster. If for
whatever reason we want to do everything with perl, this can be
requested with the "--perl" option.

Special functions
For convenience we support several special functions in any expression
passed on to awk or perl (named expressions, matches expressions,
"--eval" strings). These generally maintain some internal state, and
vnl-filter makes sure that this state is consistent. Note that these are
evaluated *after* "--skipcomments" and "--has". So any record skipped
because of a "--has" expression, for instance, will *not* be considered
in prev(), diff() and so on.

* rel(x) returns value of "x" relative to the first value of "x". For
instance we might want to see the time or position relative to the
start, not relative to some absolute beginning. Example:

$ cat tst.vnl

# time x
100 200
101 212
102 209

$ ) # read each line
{
chomp;
next unless matches; # skip non-matching lines
evalexpr();
}

--function|--sub
Evaluates the given expression as a function that can be used in other
expressions. This is most useful when you want to print something that
can't trivially be written as a simple expression. For instance:

$ cat tst.vnl
# s
1-2
3-4
5-6

$ < tst.vnl
vnl-filter --function 'before(x) { sub("-.*","",x); return x }' \
--function 'after(x) { sub(".*-","",x); return x }' \
-p 'b=before(s),a=after(s)'
# b a
1 2
3 4
5 6

See the CAVEATS section below if you're doing something
sufficiently-complicated where you need this.

--function-abs|--sub-abs
Convenience option to add an absolute-value abs() function. This is only
useful for awk programs (the default, no "--perl" given) since perl
already provides abs() by default.

--begin|--BEGIN
Evaluates the given expression in the BEGIN {} block of the generated
awk (or perl) program.

--end|--END
Evaluates the given expression in the END {} block of the generated awk
(or perl) program.

--[no]skipempty
Do [not] skip records where all fields are blank. By default we *do*
skip all empty records; to include them, pass "--noskipempty"

--skipcomments
Don't output non-legend comments

--perl
By default all procesing is performed by "mawk", but if for whatever
reason we want perl instead, pass "--perl". Both modes work, but "mawk"
is noticeably faster. "--perl" could be useful because it is more
powerful, which could be important since a number of things pass
commandline strings directly to the underlying language (named
expressions, matches expressions, "--eval" strings). Note that while
variables in perl use sigils, column references should *not* use sigils.
To print the sum of all values in column "a" you'd do this in awk

vnl-filter --eval '{suma += a} END {print suma}'

and this in perl

vnl-filter --perl --eval '{$suma += a} END {say $suma}'

The perl strings are evaluated without "use strict" or "use warnings" so
I didn't have to declare $suma in the example.

With "--perl", empty strings ("-" in the vnlog file) are converted to
"undef".

--dumpexprs
Used for debugging. This spits out all the final awk (or perl) program
we run for the given commandline options and given input. This is the
final program, with the column references resolved to numeric indices,
so one can figure out what went wrong.

--unbuffered
Flushes each line after each print. This makes sure each line is output
as soon as it is available, which is crucial for realtime output and
streaming plots.

--stream
Synonym for "--unbuffered"

CAVEATS
This tool is very lax in its input validation (on purpose). As a result,
columns with names like %CPU and "TIME+" do work (i.e. you can more or
less feed in output from "top -b"). The downside is that shooting
yourself in the foot is possible. This tradeoff is currently tuned to be
very permissive, which works well for my use cases. I'd be interested in
hearing other people's experiences. Potential pitfalls/unexpected
behaviors:

* All column names are replaced in all eval strings without regard to
context. The earlier example that reports the sum of values in a
column: vnl-filter --eval '{suma += a} END {print suma}' will work
fine if we *do* have a column named "a" and do *not* have a column
named "suma". But will not do the right thing if any of those are
violated. For instance, if a column "a" doesn't exist, then "awk"
would see "suma += a" instead of something like "suma += $5". "a"
would be an uninitialized variable, which evaluates to 0, so the
full "vnl-filter" command would not fail, but would print 0 instead.
It's the user's responsibility to make sure we're talking about the
right columns. The focus here was one-liners so hopefully nobody has
so many columns, they can't keep track of all of them in their head.
I don't see any way to resolve this without seriously impacting the
scope of the tool, so I'm leaving this alone.

* It is natural to use vnlog as a database. You can run queries with
something like

vnl-filter 'key == 5'

This works. But unlike a real database this is clearly a linear
lookup. With large data files, this would be significantly slower
than the logarithmic searches provided by a real database. The
meaning of "large" and "significant" varies, and you should test it.
In my experience vnlog "databases" scale surprisingly well. But at
some point, importing your data to something like sqlite is well
worth it.

* When substituting column names I match *either* a word-nonword
transition ("\b") *or* a whitespace-nonword transition. The word
boundaries is what would be used 99% of the time. But the keys may
have special characters in them, which don't work with "\b". This
means that whitespace becomes important: "1+%CPU" will not be parsed
as expected, which is correct since "+%CPU" is also a valid field
name. But "1+ %CPU" will be parsed correctly, so if you have weird
field names, put the whitespace into your expressions. It'll make
them more readable anyway.

* Strings passed to "-p" are split on "," *except* if the "," is
inside balanced "()". This makes it possible to say things like
vnl-filter --function 'f(a,b) { ... }' -p 'c=f(a,b)'. This is
probably the right behavior, although some questionable looking
field names become potentially impossible: "f(a" and "b)" *could*
otherwise be legal field names, but you're probably asking for
trouble if you do that.

* Currently there're two modes: a pick/print mode and an "--eval"
mode. Then there's also "--function", which adds bits of "--eval" to
the pick/print mode, but it feels maybe insufficient. I don't yet
have strong feelings about what this should become. Comments welcome

#+END_EXAMPLE

** vnl-align
#+BEGIN_EXAMPLE
NAME
vnl-align - aligns vnlog columns for easy interpretation by humans

SYNOPSIS
$ cat tst.vnl

# w x y z
-10 40 asdf -
-20 50 - 0.300000
-30 10 whoa 0.500000

$ vnl-align tst.vnl

# w x y z
-10 40 asdf -
-20 50 - 0.300000
-30 10 whoa 0.500000

DESCRIPTION
The basic usage is

vnl-align logfile

The arguments are assumed to be the vnlog files. If no arguments are
given, the input comes from STDIN.

This is very similar to "column -t", but handles "#" lines properly:

1. The first "#" line is the legend. For the purposes of alignment, the
leading "#" character and the first column label are treated as one
column

2. All other "#" lines are output verbatim.

#+END_EXAMPLE

** vnl-sort
#+BEGIN_EXAMPLE
NAME
vnl-sort - sorts an vnlog file, preserving the legend

SYNOPSIS
$ cat a.vnl
# a b
AA 11
bb 12
CC 13
dd 14
dd 123

Sort lexically by a:
$ 2. See below for details.

Past that, everything "join" does is supported, so see that man page for
detailed documentation. Note that all non-legend comments are stripped
out, since it's not obvious where they should end up.

Field names in the output
By default, the field names in the output match those in the input. This
is what you want most of the time. It is possible, however that a column
name adjustment is needed. One common use case for this is if the files
being joined have identically-named columns, which would produce
duplicate columns in the output. Example: we fixed a bug in a program,
and want to compare the results before and after the fix. The program
produces an x-y trajectory as a function of time, so both the bugged and
the bug-fixed programs produce a vnlog with a legend

# time x y

Joining this on "time" will produce a vnlog with a legend

# time x y x y

which is confusing, and *not* what you want. Instead, we invoke
"vnl-join" as

vnl-join --vnl-suffix1 _buggy --vnl-suffix2 _fixed -j time buggy.vnl fixed.vnl

And in the output we get a legend

# time x_buggy y_buggy x_fixed y_fixed

Much better.

Note that "vnl-join" provides several ways of specifying this. The above
works *only* for 2-way joins. An alternate syntax is available for N-way
joins, a comma-separated list. The same could be expressed like this:

vnl-join -a- --vnl-suffix _buggy,_fixed -j time buggy.vnl fixed.vnl

Finally, if passing in structured filenames, "vnl-join" can infer the
desired syntax from the filenames. The same as above could be expressed
even simpler:

vnl-join --vnl-autosuffix -j time buggy.vnl fixed.vnl

This works by looking at the set of passed in filenames, and stripping
out the common leading and trailing strings.

Sorting of input and output
The GNU coreutils "join" tool expects sorted columns because it can then
take only a single pass through the data. If the input isn't sorted,
then we can use normal shell substitutions to sort it:

$ vnl-join -j key <(vnl-sort -s -k key a.vnl) <(vnl-sort -s -k key b.vnl)

For convenience "vnl-join" provides a "--vnl-sort" option. This allows
the above to be equivalently expressed as

$ vnl-join -j key --vnl-sort - a.vnl b.vnl

The "-" after the "--vnl-sort" indicates that we want to sort the
*input* only. If we also want to sort the output, pass the short codes
"sort" accepts instead of the "-". For instance, to sort the input for
"join" and to sort the output numerically, in reverse, do this:

$ vnl-join -j key --vnl-sort rg a.vnl b.vnl

The reason this shorthand exists is to work around a quirk of "join".
The sort order is *assumed* by "join" to be lexicographical, without any
way to change this. For "sort", this is the default sort order, but
"sort" has many options to change the sort order, options which are
sorely missing from "join". A real-world example affected by this is the
joining of numerical data. If you have "a.vnl":

# time a
8 a
9 b
10 c

and "b.vnl":

# time b
9 d
10 e

Then you cannot use "vnl-join" directly to join the data on time:

$ vnl-join -j time a.vnl b.vnl
# time a b
join: /dev/fd/4:3: is not sorted: 10 c
join: /dev/fd/5:2: is not sorted: 10 e
9 b d
10 c e

Instead you must re-sort both files lexicographically, *and* then
(because you almost certainly want to) sort it back into numerical
order:

$ vnl-join -j time <(vnl-sort -s -k time a.vnl) <(vnl-sort -s -k time b.vnl) |
vnl-sort -s -n -k time
# time a b
9 b d
10 c e

Yuck. The shorthand described earlier makes the interface part of this
palatable:

$ vnl-join -j time --vnl-sort n a.vnl b.vnl
# time a b
9 b d
10 c e

Note that the input sort is stable: "vnl-join" will invoke "vnl-sort
-s". If you want a stable post-sort, you need to ask for it with
"--vnl-sort s...".

N-way joins
The GNU coreutils "join" tool is inherently designed to join *exactly*
two files. "vnl-join" extends this capability by chaining together a
number of "join" invocations to produce a generic N-way join. This works
exactly how you would expect with the following caveats:

* Full outer joins are supported by passing "-a-", but no other "-a"
option is supported. This is possible, but wasn't obviously worth
the trouble.

* "-v" is not supported. Again, this is possible, but wasn't obviously
worth the trouble.

* Similarly, "-o" is not supported. This is possible, but wasn't
obviously worth the trouble, especially since the desired behavior
can be obtained by post-processing with "vnl-filter".

BUGS AND CAVEATS
The underlying "sort" tool assumes lexicographic ordering, and matches
fields purely based on their textual contents. This means that for the
purposes of joining, 10, 10.0 and 1.0e1 are all considered different. If
needed, you can normalize your keys with something like this:

vnl-filter -p x='sprintf("%f",x)'

COMPATIBILITY
I use GNU/Linux-based systems exclusively, but everything has been
tested functional on FreeBSD and OSX in addition to Debian, Ubuntu and
CentOS. I can imagine there's something I missed when testing on
non-Linux systems, so please let me know if you find any issues.

ecosyste.ms

Data

Tools

Indexes

Applications

Experiments

Awesome

https://github.com/dkogan/vnlog

Awesome Lists containing this project

README