Unexpected `enomem` error during file downloading

[viralpraxis]

Hi there!

We tried to migrate from 1.18.1 to 1.20.1 and encountered increased error rate during downloading huge (~1GB) files.

I've published an MRE: https://gist.github.com/viralpraxis/6209746ee3b108473452ee810678f769
You can run it via mix run -e 'Runner.start($URI)', $URI points to a file over 1GB (serving over python -m http.server would be fine).

On 1.18.1:

21:47:47.878 [info] 2: Streaming completed
21:47:47.946 [info] 1: Streaming completed
21:47:47.990 [info] 5: Streaming completed
21:47:48.055 [info] 4: Streaming completed
21:47:48.066 [info] 3: Streaming completed

On 1.20.1:

21:48:12.494 [error] Error while streaming: enomem
21:48:12.494 [error] Error while streaming: enomem
21:48:12.494 [error] Error while streaming: enomem
21:48:12.494 [error] Error while streaming: enomem
21:48:12.494 [error] Error while streaming: enomem

In fact this error is reproducible on 1.18.2. I bisect'd this error to 5e74354 (previous commit e3872f7 works fine).

elixir 1.16.1
erlang 26.2.2

Please let my know if you need any addition information. Thanks!

[benoitc]

do you reproduce it in latest version ?

[viralpraxis]

@benoitc just checked out 1.24.1 -- this is still valid.

[zrzka]

I can confirm that downgrading to 1.18.1 fixes the enomem issue for me. In my case, it's about downloading 100 MB file from Azure Blob Storage (azurex -> httpoison -> hackney).

elixir 1.18.4 & erlang 27.3.4.1

[zrzka]

This is interesting ...

• Before we bumped Hackney to 1.24.1, we had it pinned to eca5fbb
• Then we bumped Hackney to 1.24.1, Elixir to 1.18.4 (was 1.16.2) and Erlang to 27.3.4.1 (was 26.2.2) and enomem issue appeared
• I did downgrade Hackney to 1.18.1 and the enomem issue is gone
• I bumped Hackney to eca5fbb (basically 1.20.1 + some fixes) and the issue is back

... IOW, the pattern I see here is ...

• enomem in Elixir 1.18.4 & Erlang 27.3.4.1
• no enomem in Elixir 1.16.2 & Erlang 26.2.2

[thbar]

I have also encountered this today. A couple of data points:

• The file size is 104MiB, so not huge
• we use ExAWS.S3, which leverages hackney as its default client
• for some unknown reason, it happens on my Mac, but not on our production (Ubuntu Noble) docker container (yet at least)

For comparison, we are on:

• Erlang/OTP 27 [erts-15.2.7] [source] [64-bit] [smp:16:16] [ds:16:16:10] [async-threads:1] [jit]
• Elixir 1.18.4 (compiled with Erlang/OTP 27)
• hackney 1.24.1 if I'm correct

[viralpraxis]

Alright, I took a closer look at suspicious commit 5e74354. It seems that Hackney attempts to allocate N bytes of memory when downloading a file of size N. This means that downloading a 1GB file would try to allocate 1GB of memory, which is very likely to fail.

hackney/src/hackney_response.erl

Line 369 in e2bbdf7

Transport:recv(Skt, ExpectedSize - BufSize, Timeout);

I'm fairly pretty sure it should be removed and a more reasonable buffer size (e.g., 512KB, or possibly just default zero value) should be used.

[viralpraxis]

I've opened #774. I'm not familiar with the codebase, so I'm not sure if it's correct. But at least it resolves the enomem issue (tested on 2GB downloading) and I hope it might help move us toward a valid fix more quickly.

[bszaf]

It seems that the amount of data is constrained by TCP C port driver. It forbids to ask for packets larger than 64MB:

https://github.com/erlang/otp/blob/359e254aba76c1986b671b45fd320c6cc6720ca8/erts/emulator/drivers/common/inet_drv.c#L1297
https://github.com/erlang/otp/blob/359e254aba76c1986b671b45fd320c6cc6720ca8/erts/emulator/drivers/common/inet_drv.c#L12177-L12178

[grzuy]

Seeing

%HTTPoison.Error{reason: :enomem, id: nil}

when GETting a ~75MB file from Google Storage bucket.

elixir 1.18.4
erlang 28.0.1
hackney 1.24.1

[viralpraxis]

Yeah, apparently this happens to any file >= 64MB. I believe #774 should fix that.