OS adapter and sqlsh


The OS (operating system) adapter allows you to access data in your operating system and environment using SQL queries.

It aims to solve similar problems that have traditionally been solved using UNIX command pipelines, but with the power and type-safety of SQL.

The adapter also includes a wrapper called sqlsh that allows you to execute commands from your favorite shell.

Security warning

The OS adapter launches processes, and is potentially a security loop-hole. It is included in Calcite’s “plus” module, which is not enabled by default. You must think carefully before enabling it in a security-sensitive situation.


We try to support all tables on every operating system, and to make sure that the tables have the same columns. But we rely heavily on operating system commands, and these differ widely. So:

  • These commands only work on Linux and macOS (not Windows, even with Cygwin);
  • vmstat has very different columns between Linux and macOS;
  • files and ps have the same column names but semantics differ;
  • Other commands work largely the same.

A simple example

Every bash hacker knows that to find the 3 largest files you type

$ find . -type f -print0 | xargs -0 ls -l  | sort -nr -k 5 | head -3
-rw-r--r-- 1 jhyde jhyde 194877 Jul 16 16:10 ./validate/SqlValidatorImpl.java
-rw-r--r-- 1 jhyde jhyde  73826 Jul  4 21:51 ./fun/SqlStdOperatorTable.java
-rw-r--r-- 1 jhyde jhyde  39214 Jul  4 21:51 ./type/SqlTypeUtil.java

This actually a pipeline of relational operations, each tuple represented by line of space-separated fields. What if we were able to access the list of files as a relation and use it in a SQL query? And what if we could easily execute that SQL query from the shell? This is what sqlsh does:

$ sqlsh select size, path from files where type = \'f\' order by size desc limit 3
194877 validate/SqlValidatorImpl.java
73826 fun/SqlStdOperatorTable.java
39214 type/SqlTypeUtil.java


sqlsh launches a connection to Calcite whose default schema is the OS adapter.

It uses the JAVA lexical mode, which means that unquoted table and column names remain in the case that they were written. This is consistent with how shells like bash behave.

Shell meta-characters such as *, >, <, (, and ) have to be treated with care. Often adding a back-slash will suffice.

Tables and commands

The OS adapter contains the following tables:

  • du - Disk usage (based on du command)
  • ps - Processes (based on ps command)
  • stdin - Standard input
  • files - Files (based on the find command)
  • git_commits - Git commits (based on git log)
  • vmstat - Virtual memory (based on vmstat command)

Most tables are implemented as views on top of table functions.

New data sources are straightforward to add; please contribute yours!

Example: du

How many class files, and what is their total size? In bash:

$ du -ka . | grep '\.class$' | awk '{size+=$1} END {print FNR, size}'
4416 27960

In sqlsh:

$ sqlsh select count\(\*\), sum\(size_k\) from du where path like \'%.class\'
4416 27960

The back-slashes are necessary because (, *, ), and ' are shell meta-characters.

Example: files

How many files and directories? In bash, you would use find:

$ find . -printf "%Y %p\n" | grep '/test/' | cut -d' ' -f1 | sort | uniq -c
    143 d
   1336 f

In sqlsh, use the files table:

$ sqlsh select type, count\(\*\) from files where path like \'%/test/%\' group by type
d 143
f 1336

Example: ps

Which users have processes running? In sqlsh:

$ sqlsh select distinct ps.\`user\` from ps

The ps. qualifier and back-quotes are necessary because USER is a SQL reserved word.

Now a ‘top N’ problem: Which three users have the most processes? In bash:

$ ps aux | awk '{print $1}' | sort | uniq -c | sort -nr | head -3

In sqlsh:

$ ./sqlsh select count\(\*\), ps.\`user\` from ps group by ps.\`user\` order by 1 desc limit 3
185 root
69 jhyde
2 avahi

Example: vmstat

How’s my memory?

$ ./sqlsh -o mysql select \* from vmstat
| proc_r | proc_b | mem_swpd | mem_free | mem_buff | mem_cache | swap_si | swap_so | io_bi | io_bo | system_in | system_cs | cpu_us | cpu_sy | cpu_id | cpu_wa | cpu_st |
|     12 |      0 |    54220 |  5174424 |   402180 |   4402196 |       0 |       0 |    15 |    35 |         3 |         2 |      7 |      1 |     92 |      0 |      0 |
(1 row)

Example: explain

To find out what columns a table has, use explain:

$ sqlsh explain plan with type for select \* from du
path VARCHAR CHARACTER SET "ISO-8859-1" COLLATE "ISO-8859-1$en_US$primary" NOT NULL,

Example: git

How many commits and distinct authors per year? The git_commits table is based upon the git log command.

./sqlsh select floor\(commit_timestamp to year\) as y, count\(\*\), count\(distinct author\) from git_commits group by y order by 1
2012-01-01 00:00:00 180 6
2013-01-01 00:00:00 502 13
2014-01-01 00:00:00 679 36
2015-01-01 00:00:00 470 45
2016-01-01 00:00:00 465 67
2017-01-01 00:00:00 279 53

Note that group by y is possible because sqlsh uses Calcite’s lenient mode.

Example: stdin

Print the stdin, adding a number to each line.

$ (echo cats; echo and dogs) | cat -n -
     1  cats
     2  and dogs

In sqlsh:

$ (echo cats; echo and dogs) | ./sqlsh select \* from stdin
1 cats
2 and dogs

Example: output format

The -o option controls output format.

$ ./sqlsh -o mysql select min\(size_k\), max\(size_k\) from du
| EXPR$0 | EXPR$1 |
|      0 |  94312 |
(1 row)

Format options:

  • spaced - spaces between fields (the default)
  • headers - as spaced, but with headers
  • csv - comma-separated values
  • json - JSON, one object per row
  • mysql - an aligned table, in the same format used by MySQL

Further work

The OS adapter was created in [CALCITE-1896] but is not complete.

Some ideas for further work:

  • Allow ‘-‘and ‘.’ in unquoted table names (to match typical file names)
  • Allow ordinal field references, for example ‘$3’. This would help for files that do not have named fields, for instance stdin, but you could use them even if fields have names. Also ‘$0’ to mean the whole input line.
  • Use the file adapter, e.g. select * from file.scott.emp would use the file adapter to open the file scott/emp.csv
  • More tables based on git, e.g. branches, tags, files changed in each commit
  • wc function, e.g. select path, lineCount from git_ls_files cross apply wc(path)
  • Move sqlsh command, or at least the java code underneath it, into sqlline