How to turn an ordinary gzip archive into a database

Last updated on August 25, 2020, in python

This article demonstrates how specially crafted but ordinary gzip archives can be used as a database like storage. It also introduces a Python package and explains how it works.

gzip is a popular file compression format to store large amounts of raw data. It has a good data compression ratio, but relatively slow compression/decompression speed.

Many companies use it in Big data applications when they need to store compressed CSV or JSON lines files. Such file formats are row-oriented and usually processed line by line. gzip can save a lot of space, especially when you have repetitive field names in JSON files.

Unfortunately, a compressed file can only be accessed in the streaming fashion. Even if you have a sorted file and you know that you don't need the first half of the file — you can't jump to an arbitrary point in a file without starting decompressing it from the start of the file.

gzip format

A single gzip archive can only contain one file. Any gzip file starts with a magic number (1f 8b) followed by a 10-byte header that contains metadata. After the header, it contains compressed data followed by CRC32 checksum and the size of the original file.

The most interesting detail about the format is that gzip RFC allows multiple compressed files to be concatenated into a single file. Since gzip only allows to compress one file, multiple concatenated archives will still produce one file when decompressing.

Here is how two concatenated gzip files look like:

You can think of it as concatenating two txt files. The final file will have content from both files. To combine multiple compressed files, you can use the cat command from GNU tools.

$ echo '123' | gzip > 1.txt.gz
$ echo '456' | gzip > 2.txt.gz
$ cat 1.txt.gz 2.txt.gz > 3.txt.gz
$ zcat 3.txt.gz

The zcat command is an alias to gunzip -c.

Semi-random access

An ability to concatenate multiple archives is a very handy property. By exploiting this feature and trading a little bit of speed and disk space, we can achieve semi-random access. By random access, I mean the ability to read an arbitrary offset of data from a file without reading the whole file.

To implement semi-random access, we need to split our uncompressed input into small chunks and compress them separately. We can chunk it by the number of lines. We can also chunk it by the number of bytes rounded to the nearest line end.

Let's suppose we have 1TB of JSON lines data that contains 10M of HTML webpages from Common Crawl (CC). Since it takes a lot of space, we want to compress our data. When our dataset is compressed, it only takes around 180GB, which is six times less than storing uncompressed data. HTML data have a lot of repetitive tags, and they can be compressed pretty efficiently.

We also want to have random access to be able to sample smaller subsets of data without reading the whole archive. To do so, we can chunk (partition) our input by 10-100 lines and make a gzip archive from each chunk. We can do this on the go and append all the compressed chunks on the go to a single file.

Since our gzip archive is now chunked, we can jump to a random position in a file and search for the nearest gzip block. As we already know, each gzip file starts with a magic number and a specially defined header, so it's possible to locate each chunk's start.

To be able to perform line-level random access, we can chunk our input by one line. That is, we can make a separate archive from each line of our input. This gives us better random access but comes with a cost.

An empty gzip archive takes 20 bytes of space. If your lines are very short, gzipping each line can make a compressed file even bigger than the original input.

gzip file as a database

If each row in your data has a unique key, you can index it for faster lookups. For example, each row in our Common Crawl data has a domain name as a primary key. The primary key length is relatively short compared to the content of the other fields (HTML and HTTP headers).

Since we already chunk our data, we can keep the start of the gzipped chunk for each of our primary keys. Primary keys are usually short, and we can store such information in a separate uncompressed txt file. This will be our index.

Gzipped files contain binary data. Thus, we need to keep the byte offset:


If the number of chunks is large, you can also store the line start (local character offset) inside the gzipped archive. This allows us to jump a particular line quickly. For additional speed gains, we can also keep offset size or offset end.


Every modern file-system supports random access and allows us to read an arbitrary offset of a file. Given a row from our index file, we can first jump to our specific gzip chunk. After locating it, we can decompress it on the go and skip first n characters until we hit the start of the line that we need to retrieve.

Given the index line, we can retrieve any row in almost constant time. When using an SSD storage, searching for a single line on 200GB of compressed Common Crawl data takes less than 0.1 seconds.

The plaintext index for 10 million domains takes around 500MB of disk space, which is nothing compared to the size of HTML data.

Binary search

Fast lookups sound very interesting, but searching for a specific key in the index file takes time too. For the CC use case, we need an ability to quickly find an arbitrary domain name in our index file.

The binary search algorithm can be very handy for this problem. All we need to do is to sort our index file by the key. We don't need to load the index file in memory, because we can use random access to jump to specific lines of the uncompressed file.

This article is already pretty long, so I won't get into details of how binary search works. It's a very fast algorithm, and it takes around 20 jumps on average to find the domain name in 10M of rows. If you haven't heard about it, you can have a look at a visualization.

The main idea of the algorithm is to split the available data in half on each iteration and check the key in the middle. Since all data is sorted, we can quickly locate the primary key by knowing which side to keep and split next.

When everything is combined, the complete search system takes less than 0.3 seconds to search in the 1TB of compressed data.

I came up with this idea with Michael Penkov, and we use such an approach for real tasks. We also wrote a Python package called gzipi, and the whole implementation takes less than 1k of lines when using GNU tools (such as cat, sort, and gzip). The code is easy to learn and understand.

How to use gzipi

To start using gzipi, you need to have Python and pip installed on your system:

$ pip install gzipi

Please note that you also need GNU tools on your system. It works on Linux and macOS without any problems.

After installation, you can repack your existing gzip file:

$ gzipi repack -f websites.json.gz -i index.gzi -o websites_repacked.json.gz --format json --field domain

The field argument specifies primary key. The input format could be csv or json.

Now you can search for an arbitrary key:

$ gzipi search --input-file websites_repacked.json.gz --index-file index.gzi --key

S3 support and cost savings

One of our requirements was to support S3 as a storage for our compressed files. Fortunately, S3 allows reading arbitrary offsets, so we don't need to store compressed files locally. Since binary search takes most of the time, storing index files locally significantly speeds up the search.

If you have gzipped JSON lines files, Amazon athena provides a solution to query them using SQL syntaxis. It's a very convenient solution, but Amazon charges $5 per T.B. of data scanned. Using Amazon's solution would potentially cost us thousands of dollars per month, and our solution cost us almost nothing because reading data from S3 from the same region is free.

Possible improvements

There are many ways to improve the search even more. Them main benefit of our solution is that it does not require a third-party database and takes very little RAM.

Instead of using the binary search algorithm, you can use B-tree or MARISA-trie structures. To use them, you need to store the index file in memory.

Instead of storing keys (domain names) in a file, you can also use hash tables, but that requires storing intermediate gzip files when compressing the data. When using hash tables, you don't need to store domain names. The only thing that is required is the start of the gzip chunk for each bucket.


It took me a few days to implement an initial solution to our problem, and when I was writing it, we didn't know that there are similar solutions to query billions of webpages. As it turned out later, Common Crawl and Web Archive use pretty similar solutions. Common Crawl uses WARC files to store HTTP responses, and each response is stored as a single gzip chunk. It also uses pywb library that provides an API and indexing capabilities.

Such an approach allows Common Crawl to query 2.45 billion web pages without using a database. Since all CC data is on the S3, it also scales very well. Everyone that has access to S3 can query the data without talking to a centralized server.

The main difference is that our approach is focused on JSON and CSV files that contain much smaller sets of data (narrow data) on each row. Because of that, I use tens of thousands of rows per each gzip chunk.

If you have any questions, feel free to ask them via e-mail displayed in the footer.


  • Tom 2020-09-07 #

    What about Parquet files?


    • Artem 2020-09-08 #

      It's a good format, but it does not fit my particular needs. Not every software understands it and I still need to index records for fast search.


  • Max Haeussler 2020-10-09 #

    Congratulations! You have reinvented what genomics have been doing for more than 10 years. The tools there are called bgzip and tabix and there are already python packages for it. The system is used all across genomics, for example in .vcf.gz but also for the .bam file format.


    • Aidan 2022-11-25 #

      tabix only works on chrom/pos indexing. This is more generic solution