Can you give load times for these?

> 8000128  eye5chunk.npy
> 5004297  eye5chunk_bjd_raw.jdb
>   10338  eye5chunk_bjd_zlib.jdb
>    2206  eye5chunk_bjd_lzma.jdb

For my case, I'd be curious about the time to add one 1T-entries file to another.


On 2022-08-24 20:02, Qianqian Fang wrote:

I am curious what you and other developers think about adopting JSON/binary JSON as a similarly simple, reverse-engineering-able but universally parsable array exchange format instead of designing another numpy-specific binary format.

I am interested in this topic (as well as thoughts among numpy developers) because I am currently working on a project - NeuroJSON ( - funded by the US National Institute of Health. The goal of the NeuroJSON project is to create easy-to-adopt, easy-to-extend, and preferably human-readable data formats to help disseminate and exchange neuroimaging data (and scientific data in general).

Needless to say, numpy is a key toolkit that is widely used among neuroimaging data analysis pipelines. I've seen discussions of potentially adopting npy as a standardized way to share volumetric data (as ndarrays), such as in this thread

however, several limitations were also discussed, for example

1. npy only support a single numpy array, does not support other metadata or other more complex data records (multiple arrays are only achieved via multiple files)
2. no internal (i.e. data-level) compression, only file-level compression
3. although the file is simple, it still requires a parser to read/write, and such parser is not widely available in other environments, making it mostly limited to exchange data among python programs
4. I am not entirely sure, but I suppose it does not support sparse matrices or special matrices (such as diagonal/band/symmetric etc) - I can be wrong though

In the NeuroJSON project, we primarily use JSON and binary JSON (specifically, UBJSON derived BJData format) as the underlying data exchange files. Through standardized data annotations, we are able to address most of the above limitations - the generated files are universally parsable in nearly all programming environments with existing parsers, support complex hierarchical data, compression, and can readily benefit from the large ecosystems of JSON (JSON-schema, JSONPath, JSON-LD, jq, numerous parsers, web-ready, NoSQL db ...).

I understand that simplicity is a key design spec here. I want to highlight UBJSON/BJData as a competitive alternative format. It is also designed with simplicity considered in the first place, yet, it allows to store hierarchical strongly-typed complex binary data and is easily extensible.

A UBJSON/BJData parser may not necessarily longer than a npy parser, for example, the python reader of the full spec only takes about 500 lines of codes (including comments), similarly for a JS parser

We actually did a benchmark a few months back - the test workloads are two large 2D numerical arrays (node, face to store surface mesh data), we compared parsing speed of various formats in Python, MATLAB, and JS. The uncompressed BJData (BMSHraw) reported a loading speed that is nearly as fast as reading raw binary dump; and internally compressed BJData (BMSHz) gives the best balance between small file sizes and loading speed, see our results here

I want to add two quick points to echo the features you desired in npy:

1. it is not common to use mmap in reading JSON/binary JSON files, but it is certainly possible. I recently wrote a JSON-mmap spec and a MATLAB reference implementation

2. UBJSON/BJData natively support append-able root-level records; JSON has been extensively used in data streaming with appendable nd-json or concatenated JSON (

just a quick comparison of output file sizes with a 1000x1000 unitary diagonal matrix

# python3 -m pip install jdata bjdata
import numpy as np
import jdata as jd
x = np.eye(1000);       # create a large array
y = np.vsplit(x, 5);    # split into smaller chunks'eye5chunk.npy',y);             # save npy, 'eye5chunk_bjd_raw.jdb');    # save as uncompressed bjd, 'eye5chunk_bjd_zlib.jdb', {'compression':'zlib'});  # zlib-compressed bjd, 'eye5chunk_bjd_lzma.jdb', {'compression':'lzma'});  # lzma-compressed bjd
newy=jd.load('eye5chunk_bjd_zlib.jdb'); # loading/decoding
newx = np.concatenate(newy);            # regroup chunks

here are the output file sizes in bytes:

8000128  eye5chunk.npy
5004297  eye5chunk_bjd_raw.jdb
  10338  eye5chunk_bjd_zlib.jdb
   2206  eye5chunk_bjd_lzma.jdb


On 8/24/22 15:48, Michael Siebert wrote:
Hi Matti, hi all,
@Matti: I don't know what exactly you are referring to (Pull request or the Github project, links see below). Maybe some clarification is needed, which I hereby try to do ;)
A .npy file created by some appending process is a regular .npy file and does not need to be read in chunks. Processing arrays larger than the systems memory can already be done with memory mapping (numpy.load(... mmap_mode=...)), so no third-party support is needed to do so.
The idea is not necessarily to only write some known-but-fragmented content to a .npy file in chunks or to only handle files larger than the RAM.
It is more about the ability to append to a .npy file at any time and between program runs. For example, in our case, we have a large database-like file containing all (preprocessed) images of all videos used to train a neural network. When new video data arrives, it can simply be appended to the existing .npy file. When training the neural net, the data is simply memory mapped, which happens basically instantly and does not use extra space between multiple training processes. We have tried out various fancy, advanced data formats for this task, but most of them don't provide the memory mapping feature which is very handy to keep the time required to test a code change comfortably low - rather, they have excessive parse/decompress times. Also other libraries can also be difficult to handle, see below.
The .npy array format is designed to be limited. There is a NEP for it, which summarizes the .npy features and concepts very well:
One of my favorite features (besides memory mapping perhaps) is this one:
"... Be reverse engineered. Datasets often live longer than the programs that created them. A competent developer should be able to create a solution in his preferred programming language to read most NPY files that he has been given without much documentation. ..."
This is a big disadvantage with all the fancy formats out there: they require dedicated libraries. Some of these libraries don't come with nice and free documentation (especially lacking easy-to-use/easy-to-understand code examples for the target language, e.g. C) and/or can be extremely complex, like HDF5. Yes, HDF5 has its users and is totally valid if one operates the world's largest particle accelerator, but we have spend weeks finding some C/C++ library for it which does not expose bugs and is somehow documented. We actually failed and posted a bug which was fixed a year later or so. This can ruin entire projects - fortunately not ours, but it ate up a lot of time we could have spend more meaningful. On the other hand, I don't see how e.g. zarr provides added value over .npy if one only needs the .npy features and maybe some append-data-along-one-axis feature. Yes, maybe there are some uses for two or three appendable axes, but I think having one axis to append to should cover a lot of use cases: this axis is typically time: video, audio, GPS, signal data in general, binary log data, "binary CSV" (lines in file): all of those only need one axis to append to.
The .npy format is so simple, it can be read e.g. in C in a few lines. Or accessed easily through Numpy and ctypes by pointers for high speed custom logic - not even requiring libraries besides Numpy.
Making .npy appendable is easy to implement. Yes, appending along one axis is limited as the .npy format itself. But I consider that rather to be a feature than a (actual) limitation, as it allows for fast and simple appends.
The question is if there is some support for an append-to-.npy-files-along-one-axis feature in the Numpy community and if so, about the details of the actual implementation. I made one suggestion in
and I offer to invest time to update/modify/finalize the PR. I've also created a library that can already append to .npy:
However, due to current limitations in the .npy format, the code is more complex than it could actually be (the library initializes and checks spare space in the header) and it needs to rewrite the header every time. Both could be made unnecessary with a very small addition to the .npy file format. Data would stay continuous (no fragmentation!), there should just be a way to indicate that the actual shape of the array should derived from the file size.
Best, Michael

On 24. Aug 2022, at 19:16, Matti Picus <> wrote:
Sorry for the late reply. Adding a new "*.npy" format feature to allow writing to the file in chunks is nice but seems a bit limited. As I understand the proposal, reading the file back can only be done in the chunks that were originally written. I think other libraries like zar or h5py have solved this problem in a more flexible way. Is there a reason you cannot use a third-party library to solve this? I would think if you have an array too large to write in one chunk you will need third-party support to process it anyway.


NumPy-Discussion mailing list --
To unsubscribe send an email to
Member address:

NumPy-Discussion mailing list --
To unsubscribe send an email to
Member address:

NumPy-Discussion mailing list --
To unsubscribe send an email to
Member address: