How to make a dumthe application reads gzip files (on-the-fly) without code changes

gzip stores some of the original file's metadata in record header, including the file modification time and filename, if available. See GZIP file format specification.

So it's expected that your two gzip files aren't identical. You can work around this by passing gzip the -n flag, which stops it from including the original filename and timestamp in the header.

This does NOT work:

# echo Hello World > example.txt
# gzip < example.txt > example.txt # WRONG!
# file example.txt
example.txt: gzip compressed data, from Unix, last modified: Thu Mar 21 19:45:29 2013
# gunzip < example.txt
<empty file>

This is a race condition:

# echo Hello World > example.txt
# dd if=example.txt | gzip | dd of=example.txt # still WRONG!
# gunzip < example.txt 
Hello World # may also be empty

The problem is that the > example.txt (or dd of=example.txt for that matter) kills the file before the other process has the chance to read it. So there is no obvious solution, which is why you should stick to mv.

There are a number of ways you could cheat. You can open the file, then unlink it - the file will continue to exist until you close it - and then create a new file with the same name and write the gzipped data to that. However I do not know an obvious way to coerce bash to use that, and even if I did, my answer would still be:

Don't even do it.

If gzip fails for any reason, or any problem occurs, like you running out of space while gzipping (because other processes are writing, or gzip result is larger than the input - which happens for random data - etc.), you just lost your file. Congratulations!

Create a separate file and mv on success. That's the simplest, easy to understand, and most reliable method you will ever find.