mxnet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Skalicky, Sam" <>
Subject Re: Stopping nightly releases to Pypi
Date Sun, 08 Dec 2019 20:48:51 GMT
Thanks Sheng,

Looks like 12/8 builds are working as expected too:

(Plain mxnet, no mkl no cuda)

I’ll try and get this updated on the installation docs page tomorrow.


On Dec 7, 2019, at 5:30 PM, Sheng Zha <<>>

Heres a set of links for today’s builds

(Plain mxnet, no mkl no cuda)

These links are not utilizing the s3 accelerate feature (i.e. not backed by cloudfront edges).
Please use<> instead. The updated links are:
(Plain mxnet, no mkl no cuda)

When updating the installation doc we should use<>
domain name too.


On 2019/12/07 17:39:40, "Skalicky, Sam" <<>>
Hi MXNet Community,

We have been working on getting nightly builds fixed and made available again. We’ve made
another system using AWS CodeBuild & S3 to work around the problems with Jenkins CI, PyPI,
etc. It is currently building all the flavors and publishing to an S3 bucket here:

There are folders for each set of nightly builds, try out the wheels starting today 2019-12-07.
Builds start at 1:30am PT (9:30am GMT) and arrive in the bucket 30min-2hours later. Inside
each folder are the wheels for each flavor of MXNet. Currently we’re only building for linux,
builds for windows/Mac will come later.

If you want to download the wheels easily you can use a URL in the form of:<YYYY-MM-DD>/dist/<mxnet_build>-1.6.0b<YYYYMMDD>-py2.py3-none-manylinux1_x86_64.whl

Heres a set of links for today’s builds

(Plain mxnet, no mkl no cuda)

You can easily install these pip wheels in your system either by downloading them to your
machine first and then installing by doing:

pip install /path/to/downloaded/wheel.whl

Or you can install directly by just giving the link to pip like this:

pip install

Credit goes to everyone involved (in no particular order)
Rakesh Vasudevan
Zach Kimberg
Manu Seth
Sheng Zha
Jun Wu
Pedro Larroy
Chaitanya Bapat


On Dec 5, 2019, at 1:16 AM, Lausen, Leonard <<><>>

We don't loose pip by hosting on S3. We just don't host nightly releases on Pypi
servers and mirror them to several hundred mirrors immediately after each build
is published which is very expensive for the Pypi project.. People can still
install the nightly builds with pip by specifying the -f option.

Uploading weekly releases to Pypi will reduce the cost for Pypi by ~75% [1]. It
may be acceptable to Pypi, but does it make sense for us? I'm not convinced
weekly release on Pypi is a good idea. Consider one release is buggy, users will
need to wait for 7 days for a fix. It doesn't provide good user experience.
If someone has a stronger conviction about the value of weekly releases on Pypi,
that person shall please go ahead and propose it in a separate discussion

Currently we don't have generally working nightly builds to Pypi and as a matter
of fact we know that we can't have them due to Pypi's policy and our apparent
need for large binaries. Given this fact and that no objection was raised by
2019-12-05 at 05:42 UTC, I conclude we have lazy consensus on stopping upload
attempts of nightly builds to Pypi.

With consensus established, we can change the CI job to stop trying to upload
the nightly builds and then request Pypi to increase the limit. Then we have one
less blocker for the 1.6 release.

Best regards

[1]: Lower cost due to less releases, but higher cost due to 500MB -> 800MB
limit increase. Assuming that the limit increase translates into actually larger

On Wed, 2019-12-04 at 22:20 +0100, Marco de Abreu wrote:
Are weekly releases an option? It was brought up as concern that we might
lose pip as a pretty common distribution channel where people consume
nightly builds. I don't feel like that concern has been properly addressed
so far.


Lausen, Leonard <<><>>
schrieb am Mi., 4. Dez. 2019,

As a simple POC to test distribution, you can try installing MXNet based on
these 3 URLs:

pip install --no-cache-dir
pip install --no-cache-dir
pip install --no-cache-dir

where --no-cache-dir prevents caching the downloaded file, for the purpose
testing. (cu101 chosen based on large size)

The first URL uses standard S3 bucket in US. The second uses S3 Accelerate
on CloudFront CDN. And the third uses CloudFront CDN. I'm adding the third
as S3 Accelerate may or may not use all new CloudFront endpoints yet.

Regarding voting: Uploading to Pypi is currently impossible, which is a
(so there is no option to continue as we do currently). Pypi folks
they will unblock our uploads to Pypi once we stop uploading nightly
and taking up 20% of their ressources [1].

If there are any shortcomings or problems identified with uploading to S3,
can work to address them. But for now, status quo is broken and this seems
only solution addressing Pypi's problem.

I don't mind if you state that you object to lazy consensus and start a
vote. If
your "maybe [...] start a proper vote" was supposed to be an objection to
consensus, please state so clearly (I'm not sure if "maybe" qualifies as
objection). Though I think it only makes sense with at least 2 options to
on. Status quo is not a meaningful option, as it is already broken.

Best regards


On Tue, 2019-12-03 at 19:28 +0100, Marco de Abreu wrote:
Excellent! Could we maybe come up with a POC and a quick writeup and then
start a proper vote after everyone verified that it covers their

Sheng Zha <> schrieb am Di., 3. Dez. 2019, 19:24:

Yes, there is. We can also make it easier to access by using a
geo-location based DNS server so that China users are directed to that
local mirror. The rest of the world is already covered by the global


On 2019/12/03 18:22:22, Marco de Abreu <>
Isn't there an s3 endpoint in Beijing?

It seems like this topic still warrants some discussion and thus I'd

if we don't move forward with lazy consensus.


Tao Lv <> schrieb am Di., 3. Dez. 2019, 14:31:

* For pypi, we can use mirrors.

On Tue, Dec 3, 2019 at 9:28 PM Tao Lv <> wrote:

As we have many users in China, I'm considering the
accessibility of
For pip, we can mirrors.

On Tue, Dec 3, 2019 at 3:24 PM Lausen, Leonard


I would like to remind everyone that lazy consensus is assumed
if no
are raised before 2019-12-05 at 05:42 UTC. There has been some

the proposal, but to my understanding no objections were
If the proposal is accepted, MXNet releases would be installed
 pip install mxnet

And release candidates via

pip install --pre mxnet

(or with the respective cuda version specifier appended etc.)

To obtain releases built automatically from the master branch,
would need
to specify something like "-f" option to
Best regards

On Mon, 2019-12-02 at 05:42 +0000, Lausen, Leonard wrote:
Hi MXNet Community,

since more than 2 months our binary Python nightly releases

on Pypi
are broken. The problem is that our binaries exceed Pypi's
Decreasing the binary size by adding compression breaks

Sheng requested Pypi to increase their size limit:

Currently "the biggest cost for PyPI from [the many MXNet
release to Pypi] is the bandwidth consumed when several
to mirror each release immediately after it's published". So
inclined to allow us to upload even larger binaries on a
Their compromise is to allow it on a weekly cadence.

However, I would like the community to revisit the necessity
releasing pre-
release binaries to Pypi on a nightly (or weekly) cadence.

Instead, we
release nightly binaries ONLY to a public S3 bucket and
install from there. On our side, we only need to prepare a
document that
contains links to all released nightly binaries.
Finally users will install the nightly releases via

pip install --pre mxnet-cu101 -f

Instead of

pip install --pre mxnet-cu101

Of course proper releases and release candidates should
still be
via Pypi. Thus releases would be installed via

pip install mxnet-cu101

And release candidates via

pip install --pre mxnet-cu101

This will substantially reduce the costs of the Pypi project
the installation experience provided by PyTorch. I don't
think the
benefit of
not including "-f"
matches the costs we currently externalize to the Pypi team.

This suggestion seems uncontroversial to me. Thus I would
like to
consensus. If there are no objections, I will assume lazy

consensus on
nightly releases to Pypi in 72hrs.

Best regards

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message