Log rotating with a Bash script

Question

I have the following issue:

I have an application, which continuously produces output to stderr and stdout. The output of this application is captured in a logfile (the app is redirected as: &> log.txt ). I don't have any options to produce a proper logging to file for this.

Now, I have a cron job, which runs every hour and beside of doing other things, it also tries to rotate this logfile above, by copying it to log.txt.1 and then creates an empty file and copies it to log.txt

It looks like:

cp log.txt log.txt.1
touch /tmp/empty
cp /tmp/empty log.txt

The problem is, that the application is still writing to it, and because of this I get some very strange stuff in the log.txt.1, it starts with a lot of garbage characters, and the actual log file is somewhere at the end.

Do you have any idea, how to make a correct log rotating for this specific situation (I also tried cat log.txt > log.txt.1, does not work)? Using logrotate for this specific application not an option, there is a whole mechanism behind the scenes that I may not change.

Thanks, f.

Answer 1

Okay, here's an idea, inspired by http://en.wikibooks.org/wiki/Bourne_Shell_Scripting/Files_and_streams

1. make a named pipe:

   mkfifo /dev/mypipe
   

2. redirect stdout and stderr to the named pipe:

   &> /dev/mypipe
   

3. read from mypipe into a file:

   cat < /dev/mypipe > /var/log/log.txt &
   

4. when you need to log-rotate, kill the cat, rotate the log, and restart the cat.

Now, I haven't tested this. Tell us how it goes.

Note: you can give the named pipe any name, like /var/tmp/pipe1 , /var/log/pipe , /tmp/abracadabra , and so on. Just make sure to re-create the pipe after booting before your logging-script runs.

---

Alternatively, don't use cat, but use a simple script file:

#!/bin/bash

while : ; do
  read line
  printf "%s\n" "$line"
done

This script guarantees an output for every newline read. (cat might not start outputting until its buffer is full or it encounters an EOF)

---

Final -- and TESTED -- attempt

IMPORTANT NOTE: Please read the comments from @andrew below. There are several situations which you need to be aware of.

Alright! Finally got access to my Linux box. Here's how:

Step 1: Make this recorder script:

#!/bin/bash

LOGFILE="/path/to/log/file"
SEMAPHORE="/path/to/log/file.semaphore"

while : ; do
  read line
  while [[ -f $SEMAPHORE ]]; do
    sleep 1s
  done
  printf "%s\n" "$line" >> $LOGFILE
done

Step 2: put the recorder into work:

1. Make a named pipe:

   mkfifo $PIPENAME
   

2. Redirect your application's STDOUT & STDERR to the named pipe:

   ...things... &> $PIPENAME
   

3. Start the recorder:

   /path/to/recorder.sh < $PIPENAME &
   

   You might want to nohup the above to make it survive logouts.

4. Done!

Step 3: If you need to logrotate, pause the recorder:

touch /path/to/log/file.semaphore
mv /path/to/log/file /path/to/archive/of/log/file
rm /path/to/log/file.semaphore

I suggest putting the above steps into its own script. Feel free to change the 2nd line to whatever log-rotating method you want to use.

---

Note : If you're handy with C programming, you might want to make a short C program to perform the function of recorder.sh. Compiled C programs will certainly be lighter than a nohup-ed detached bash script.

---

Note 2: David Newcomb provided a helpful warning in the comments: While the recorder is not running then writes to the pipe will block and may cause the program to fail unpredictably. Make sure the recorder is down (or rotating) for as short time as possible.

So, if you can ensure that rotating happens really quickly, you can replace sleep (a built-in command which accepts only integer values) with /bin/sleep (a program that accepts float values) and set the sleep period to 0.5 or shorter.

Answer 2

First of all, you really should not reinvent the square wheel here. Your peers are probably against rotating the logs on daily schedule which automatically applies to all scripts in /etc/logrotate.d/ - this can be avoided by placing the script elsewhere.

---

Now, the standard approach to log rotation (that is implemented in logrotate) can be implemented by any other facility just as well. E.g. here's a sample implementation in bash:

MAXLOG=<maximum index of a log copy>
for i in `seq $((MAXLOG-1)) -1 1`; do
    mv "log."{$i,$((i+1))}    #will need to ignore file not found errors here
done 
mv log log.1    # since a file descriptor is linked to an inode rather than path,
                #if you move (or even remove) an open file, the program will continue
                #to write into it as if nothing happened
                #see https://stackoverflow.com/questions/5219896/how-do-the-unix-commands-mv-and-rm-work-with-open-files
<make the daemon reopen the log file with the old path>

The last item is done by sending SIGHUP or (less often) SIGUSR1 and having a signal handler in the daemon that replaces the corresponding file descriptor or variable. This way, the switch is atomic, so there's no interruption in logging availability. In bash, this would look like:

trap { exec &>"$LOGFILE"; } HUP

---

The other approach is to make the writing program itself keep track of the log size each time it writes to it and do the rotation. This limits your options in where you can write to and what rotation logic is to what the program itself supports. But it has the benefit of being a self-contained solution and checking the log size at each write rather than on schedule. Many languages' standard libraries have such a facility. As a drop-in solution, this is implemented in Apache's rotatelogs:

<your_program> 2>&1 | rotatelogs <opts> <logfile> <rotation_criteria>

Answer 3

I wrote a logrotee this weekend. I probably wouldn't if I've read @JdeBP's great answer about multilog before.

I focused on it being lightweight and being able to bzip2 its output chunks like:

verbosecommand | logrotee  \
  --compress "bzip2 {}" --compress-suffix .bz2 \
  /var/log/verbosecommand.log

There's a lot of to be done and tested yet, though.

Answer 4

You can also pipe your output thru Apache rotatelogs utility. Or following script:

#!/bin/ksh
#rotatelogs.sh -n numberOfFiles pathToLog fileSize[B|K|M|G]

numberOfFiles=10

while getopts "n:fltvecp:L:" opt; do
    case $opt in
  n) numberOfFiles="$OPTARG"
    if ! printf '%s\n' "$numberOfFiles" | grep '^[0-9][0-9]*$' >/dev/null; then
      printf 'Numeric numberOfFiles required %s. rotatelogs.sh -n numberOfFiles pathToLog fileSize[B|K|M|G]\n' "$numberOfFiles" 1>&2
      exit 1
    elif [ $numberOfFiles -lt 3 ]; then
      printf 'numberOfFiles < 3 %s. rotatelogs.sh -n numberOfFiles pathToLog fileSize[B|K|M|G]\n' "$numberOfFiles" 1>&2
    fi
  ;;
  *) printf '-%s ignored. rotatelogs.sh -n numberOfFiles pathToLog fileSize[B|K|M|G]\n' "$opt" 1>&2
  ;;
  esac
done
shift $(( $OPTIND - 1 ))

pathToLog="$1"
fileSize="$2"

if ! printf '%s\n' "$fileSize" | grep '^[0-9][0-9]*[BKMG]$' >/dev/null; then
  printf 'Numeric fileSize followed by B|K|M|G required %s. rotatelogs.sh -n numberOfFiles pathToLog fileSize[B|K|M|G]\n' "$fileSize" 1>&2
  exit 1
fi

sizeQualifier=`printf "%s\n" "$fileSize" | sed "s%^[0-9][0-9]*\([BKMG]\)$%\1%"`

multip=1
case $sizeQualifier in
B) multip=1 ;;
K) multip=1024 ;;
M) multip=1048576 ;;
G) multip=1073741824 ;;
esac

fileSize=`printf "%s\n" "$fileSize" | sed "s%^\([0-9][0-9]*\)[BKMG]$%\1%"`
fileSize=$(( $fileSize * $multip ))
fileSize=$(( $fileSize / 1024 ))

if [ $fileSize -le 10 ]; then
  printf 'fileSize %sKB < 10KB. rotatelogs.sh -n numberOfFiles pathToLog fileSize[B|K|M|G]\n' "$fileSize" 1>&2
  exit 1
fi

if ! touch "$pathToLog"; then
  printf 'Could not write to log file %s. rotatelogs.sh -n numberOfFiles pathToLog fileSize[B|K|M|G]\n' "$pathToLog" 1>&2
  exit 1
fi

lineCnt=0
while read line
do
  printf "%s\n" "$line" >>"$pathToLog"
  lineCnt=$(( $lineCnt + 1 ))
  if [ $lineCnt -gt 200 ]; then
    lineCnt=0
    curFileSize=`du -k "$pathToLog" | sed -e 's/^[  ][  ]*//' -e 's%[   ][  ]*$%%' -e 's/[  ][  ]*/[    ]/g' | cut -f1 -d" "`
    if [ $curFileSize -gt $fileSize ]; then
      DATE=`date +%Y%m%d_%H%M%S`
      cat "$pathToLog" | gzip -c >"${pathToLog}.${DATE}".gz && cat /dev/null >"$pathToLog"
      curNumberOfFiles=`ls "$pathToLog".[0-9][0-9][0-9][0-9][0-9][0-9][0-9][0-9]_[0-9][0-9][0-9][0-9][0-9][0-9].gz | wc -l | sed -e 's/^[   ][  ]*//' -e 's%[   ][  ]*$%%' -e 's/[  ][  ]*/[    ]/g'`
      while [ $curNumberOfFiles -ge $numberOfFiles ]; do
        fileToRemove=`ls "$pathToLog".[0-9][0-9][0-9][0-9][0-9][0-9][0-9][0-9]_[0-9][0-9][0-9][0-9][0-9][0-9].gz | head -1`
        if [ -f "$fileToRemove" ]; then
          rm -f "$fileToRemove"
          curNumberOfFiles=`ls "$pathToLog".[0-9][0-9][0-9][0-9][0-9][0-9][0-9][0-9]_[0-9][0-9][0-9][0-9][0-9][0-9].gz | wc -l | sed -e 's/^[   ][  ]*//' -e 's%[   ][  ]*$%%' -e 's/[  ][  ]*/[    ]/g'`
        else
          break
        fi
      done
    fi
  fi
done

Answer 5

You can leverage rotatelogs (docs here). This utility will decouple your script's stdout from the log file, managing the rotation in a transparent way. For example:

your_script.sh | rotatelogs /var/log/your.log 100M

will automatically rotate the output file when it reaches 100M (can be configured to rotate based on a time interval).

Answer 6

A simplest script could be something like this rotatelog.sh:

#! /bin/bash

#DATE_FMT="%Y%m%d-%H%M" # for testing
#DATE_FMT="%Y%m%d-%H"   # rotate each hour
DATE_FMT="%Y%m%d"       # rotate each day

if [ "$1" != "" ]
then
        f=$1
else
        f="rotatelog"
fi

p=$(date +${DATE_FMT})

r=$f-$p.log
exec 2>&1 > $r

while read l
do
        d=$(date +${DATE_FMT})
        if [ $p != $d ]
        then
                x=$r
                p=$d
                r=$f-$p.log
                exec 2>&1 > $r
                gzip $x
        fi
        echo $l
done

You can use like:

your_process | rotatelog.sh yout_log_path_pattern