On handy docker images

Motivated by successful stripping problematic dependencies from Redundans, I have decided to generate smaller Docker image, starting with Alpine Linux image (2Mb / 5Mb after downloading) instead of Ubuntu (49Mb / 122Mb). Previously, I couldn’t really rely on Alpine Linux, because it was impossible to make these problematic dependencies running… But now it’s whole new world of possibilities 😉

There are very few dependencies left, so I have started… (You can find all the commands below).

  1. First, I have check what can be installed from package manager.
    Only Python and Perl.

  2. Then I have checked if any of binaries are working.
    For example, GapCloser is provided as binary. It took me some time to find source code…
    Anyway, none of the binaries worked out of the box. It was expected, as Alpine Linux is super stripped…

  3. I have installed build-base in order to be able to build things.
    Additionally, BWA need zlib-dev.

  4. Alpine Linux doesn’t use standard glibc library, but musl-libc (you can read more about differences between the two), so some programmes (ie. BWA) may be quite reluctant to compile.
    After some hours of trying & thanks to the help of mp15, I have found a solution, not so complicated 🙂

  5. I have realised, that Dockerfile doesn’t like standard BASH brace expansion, that is working otherwise in Docker Alpine console…
    so ls *.{c,h} should be ls *.c *.h

  6. After that, LAST and GapCloser compilation were easy, relatively 😉

Below, you can find the code from Docker file (without RUN commands).

apk add --update --no-cache python perl bash wget build-base zlib-dev
mkdir -p /root/src && cd /root/src && wget http://downloads.sourceforge.net/project/bio-bwa/bwa-0.7.15.tar.bz2 && tar xpfj bwa-0.7.15.tar.bz2 && ln -s bwa-0.7.15 bwa && cd bwa && \
cp kthread.c kthread.c.org && echo "#include <stdint.h>" > kthread.c && cat kthread.c.org >> kthread.c && \
sed -ibak 's/u_int32_t/uint32_t/g' `grep -l u_int32_t *.c *.h` && make && cp bwa /bin/ && \
cd /root/src && wget http://liquidtelecom.dl.sourceforge.net/project/soapdenovo2/GapCloser/src/r6/GapCloser-src-v1.12-r6.tgz && tar xpfz GapCloser-src-v1.12-r6.tgz && ln -s v1.12-r6/ GapCloser && cd GapCloser && make && cp bin/GapCloser /bin/ && \
cd /root/src && wget http://last.cbrc.jp/last-744.zip && unzip last-744.zip && ln -s last-744 last && cd last && make && make install && \
cd /root/src && rm -r last* bwa* GapCloser* v* 

# SSPACE && redundans in /root/srt
cd /root/src && wget -q http://www.baseclear.com/base/download/41SSPACE-STANDARD-3.0_linux-x86_64.tar.gz && tar xpfz 41SSPACE-STANDARD-3.0_linux-x86_64.tar.gz && ln -s SSPACE-STANDARD-3.0_linux-x86_64 SSPACE && wget -O- -q http://cpansearch.perl.org/src/GBARR/perl5.005_03/lib/getopts.pl > SSPACE/dotlib/getopts.pl && \
wget --no-check-certificate -q -O redundans.tgz https://github.com/lpryszcz/redundans/archive/master.tar.gz && tar xpfz redundans.tgz && mv redundans-master redundans && ln -s /root/src/redundans /redundans && rm *gz

apk del wget build-base zlib-dev 
apk add libstdc++

After building & pushing, I have noticed that Alpine-based image is slightly smaller (99Mb), than the one based on Ubuntu (127Mb). Surprisingly, Alpine-based image is larger (273Mb) than Ubuntu-based (244Mb) after downloading. So, I’m afraid all of these hours didn’t really bring any substantial reduction in the image size.

I was very motivated to build my application on Alpine Linux and expected substantial size reduction. But I’d say that relying on Alpine Linux image doesn’t always pay off in terms of smaller image size, forget about production time… And this I know from my own experience.
But maybe I didn’t something wrong? I’d be really glad for some advices/comments!

Nevertheless, stripping a few dependencies from my application (namely Biopython, numpy & scipy), resulted in much more compact image even using Ubuntu-based image (127Mb vs 191Mb; and 244Mb vs 440Mb after downloading). So I think this is the way to go 🙂

On simplifying dependencies

Lately, to make Redundans more user friendly, I have simplified it’s dependencies, by replacing Biopython, numpy, scipy and SQLite with some (relatively) simple functions or modules.

Here, I will just focus on replacing Biopython, particularly SeqIO.index_db with FastaIndex. You may ask yourself, why I have invested time in reinventing the wheel. I’m big fan of Biopython, yet it’s huge project and some solutions are not optimal or require problematic dependencies. This is the case with SeqIO.db_index, that relies on SQLite3. Here again, I’m a big fan of SQLite, yet building Biopython with SQLite enabled proved not to be very straightforward for non-standard systems or less experience users. Beside, on some NFS settings, the SQLite3 db cannot be created at all.

Ok, let’s start from the basics. SeqIO.index_db allows random access to sequence files, so for example you can rapidly retrieve any entry from very large file. This is achieved by storing the ID and position of each entry from particular file in database, SQLite3 db. Then, if you want to retrieve particular record, SeqIO.index_db looks up if this record is present in SQLite3 db, retrieves record position in the file and reads only small chunk of this file instead of parsing entire file every time you want to get some record(s).
Similar feature is offered by samtools faidx, but in this case, the coordinates of each entry are stored in tab-delimited file .fai (more info about .fai). This format can be easily read & write by any programme, so I have decided to use it. In addition, I have realised, that samtools faidx is flexible enough, so you can add additional columns to the .fai without interrupting its functionality, but about that later…

In Redundans, I’ve been using SeqIO.index_db during assembly reduction (fasta2homozygous.py). Additionally, beside storing index, I’ve been also generating statistics for every FastA file, like number of contigs, cumulative size, N50, N90, GC and so on. I have realised, that these two can be easily combined, by extending .fai with four additional columns, storing number of occurencies for A, C, G & T in every sequence. Such .fai is compatible with samtools faidx and provides very easy way of calculating bunch of statistics about this file.
All of these, I’ve implemented in FastaIndex. Beside being dependency-free & very handy indexer, it can be used also as alternative to samtools faidx to retrieve sequences from large FastA files.

# retrieve bases from 20 to 60 from NODE_2
./FastaIndex.py -i test/run1/contigs.fa -r NODE_2_length_7674_cov_46.7841_ID_3:20-60
#Time elapsed: 0:00:00.014243

samtools faidx test/run1/contigs.fa NODE_2_length_7674_cov_46.7841_ID_3:20-60