HDF5 and BioSQL solutions

Question

I'm looking at better database/storage solutions for NCBI virus data, with all attributes particularly year and country of isolation, together with structural data, possible antibody data, T-cell data and bioinformatics data such as %GC etc... The amount of data is comparatively small, with a maximum size of 3 x ([6000 < no. viruses < 20000] for 10 genes) and each entry will contain nucleotide, amino acid and one or more pieces of alignment data (i.e. with indels).

The data will be used for numerous different calculations from structural biology, phylogenetics, and deep learning. Some of the calculations will require parallel processing. A typical calculation would be to identify a clade/similarity/difference then screen the database to look at the isolates information.

Previously I've used lots of flat files which are loaded at initiation and are held within a data pipeline either as an object (Python) or especially a multi-dimensional hash (Perl, preferred).

The two solutions that I think look suitable are:

1. BioSQL - The official Genbank solution, however my SQL commands are very basic and the documentation appears minimal (for BioSQL)
2. HDF5 (hierarchical data format) - this looks good because it uses complex hashes (dictionaries), and allows layering of multiple information onto a single Genbank sequence code (key). It has strong documentation. However, I understand this isn't a "database" (long term storage), but thats how I intend to use it.

JSON/Pickling I had a look at JSON, which to me looks okay, because I'm fond of complex hashes / dictionaries, but it appears to be more of a Web solution. Finally, "pickling" (Python) is cool, very cool, but I understand this not a database solution, because it risk compatibility issues, e.g. with future Python releases.

I am interested in knowing what archival systems board members are using, and a critique whether my assessment is ok for moving forward.

Answer 1

Just to conclude, @user172818 and @AlexReynolds are collectively saying either learn SQL or stick with CSV.

---

Update, I can provide an answer. It's HDF5 with pandas dataframes (Python).

Simply

df.to_hdf('path/file.h5', 'key', format='table')

This isn't SQL but it works:

What it gives,

• it's fast, that's it's purpose (not my main reason though)
• it perfectly preserves all pandas dataframe information, this is important
• The 'key' ( object) allows you to store multiple dataframes into the same file, you just provide different key names

The last point is very useful, because you can pull out a dataframe of choice and the SQL commands are mimicked within pandas

I ran into trouble using CSV to store/recover complex pandas dataframe formats.

So it's good enough and much faster than SQL. I could have problems with format changes in time (unlike SQL), but for the here and now it works well.

---

Just to add a caveat, for dictionaries the json dump works well, but it is just shunting dictionaries around, so isn't relevant.

Answer 2

Just to conclude on the solution to this ancestral question. HDF5 is very useful, but there are better solutions for this specific question if it is focused on pandas dataframes. If you've lots of data the pythonic way to do this is pandas and therefore storing as a pandas dataframe makes sense.

For long term storage, which was the original question, the storage of choice for pandas is parquet. Described here

Apache Parquet provides a partitioned binary columnar serialization for data frames. It is designed to make reading and writing data frames efficient, and to make sharing data across data analysis languages easy. Parquet can use a variety of compression techniques to shrink the file size as much as possible while still maintaining good read performance.

In the following example, two methods parquet methods are used for a pandas dataframe proteinDF which is first written and then imported.

from fastparquet import write
write('outfile.parq', proteinDF)

and

import pandas as pd
pd.read_parquet('proteindata.parquet', engine='fastparquet')

The above approach appears to have received most current developmental attention. However,

import pyarrow as pa
import pyarrow.parquet as pq
arrow_table = pa.Table.from_pandas(proteinDF)
pq.write_table(arrow_table, 'myoutpath', use_dictionary=False,
                   compression='SNAPPY')

and

import pandas as pd
pd.read_parquet('example_pa.parquet', engine='pyarrow')

The above approach is compatible with numba. 'Multithreading' is further described here.

The differences between the methods appear slight. Parallelisation is a big deal and depends on the problem.

A general discussion on the methods is here

HDF5 vs parquet The backdrop for parquet is the method is used for batch I/O not record by record writing to disk and that is this question and for that is pretty fast (at a guess bit better than pickle, but a lot more compression). HDF5 is perfect for rapid record by record writing and that is what it is designed for, so 'benchmarks' comparing I/O speed are not realistic unless they compare batch vs. record by record timings, where HDF5 excels. parquet does achieve significantly better data compression and is considered a long term storage solution. In my hands HDF5 for fasta writing results in around half the size of the flat file (not much).