warm walrus
Bug Report:
Steps to reproduce:
- Go into any conversation
- Write anything in the chat box an send
Expected result:
A full response with no errors
Actual result:
It responds but at the end of the response it says "network error" and you cannot continue the response
Additional information
Browser: chrome
OS: windows 10 android 13
grizzled dock
Same here
wheat anvil
Anyone able to capture this error in Chrome/Firefox's network tools? That could help me figure out what's going on
Steps:
high basalt
i haven't been able to detect it happen, because for me, it only happens when the message is done loading so idk how the websocket is closing
but this is happening for me as well
however it only happens when GPT uses analysis
nothing in it worth noting
wheat anvil
hmph, looks like Chrome has some bug or something where it doesn't show what happens or what is returned by that request...
high basalt
let me try firefox
wheat anvil
You are getting the "network error" message though?
high basalt
yes the message terminates and it says failed to analyze, and offers the option to regenerate the message (which works but just fails again when "done" analyzing)
wheat anvil
What does it say inside the red box?
high basalt
nothing i can see
wheat anvil
Oh, that's a different error from the "network error" reported in this thread. Can you start a new thread and also try asking ChatGPT what error it ran into?
high basalt
okay sure
lone thorn
I have been experiencing lately while using ChatGPT. The problem primarily involves slow response times and frequent network errors.
It has become quite frustrating, as I rely on ChatGPT for various tasks, and these issues are hampering my productivity. I kindly request your assistance in investigating and resolving this matter.
warm walrus
I give up on trying to do anything I'm so done at this point this has happened since the beginning of this month none of any moderators or the bug fixing team has been helpful at all I'm just going to use a different ai since chatgpt is so buggy and trashy
wheat anvil
Something came to mind - do you all see any pattern in how much time passes between when you submit your prompt and when you get the error? Like if it's always the same amount of time or not.
serene pasture
I get this all the time, trying to pin it down but it comes and goes. Not sure what is going on, trying my fiber connection and will move to my phone hotspot to see if I can rule something out. I also have two other cellular based internet backups, and will test with those.
Glad to see a bug report for this tho, it's been really terrible!
One thing I noticed is that when this happens, I can leave it and go to my phone and the entire response is there waiting for me on the chat!
and frozen... first on the last "Gen-Z" header, then now here
On my phone it's complete
I do have an adblocker, but a while ago I tuned it off to test and it still happened (this may be the refused connections to stats collection services, etc)
the conversations endpoint just seems to stall while streaming, basically randomly
this time is just hung for 20s or so, then continued for a sentence, and hung again
serene pasture
during a complete stall, not yet a network error. marked packet is when it started
and there's the network error
seems like this is all HTTP/3 wonder if that's a thread to pull on
serene pasture
haven't had a single network error since disabling HTTP/3 aka QUIC in Brave/Chrome
when this happens and a connection stalls on QUIC, even refreshing the tab fails, or closing the tab and opening a new one. browser gets quite stuck
serene pasture
things are very weird, works fine on IPv4 with QUIC and without, IPv6 without QUIC is fine, but IPv6 with QUIC has the issue
I'm running Mikrotik routers, on AT&T UVerse... will resume tomorrow debugging
it happens more on GPT4 for me for some reason, maybe because the type of responses, but now I hit the limit too
someone noted the same thing with QUIC on in Brave
serene pasture
for now I've blocked outbound UDP 80/443 on my router, and that forces everything away from QUIC. I also have a totally weird thing that happens on MS Teams (moved from Zoom) at work, where it stutters every few minutes. I wonder if this is related as Teams uses QUIC now. will report back tomorrow
you can also disable QUIC at brave://flags/#enable-quic or chrome://flags/#enable-quic
wheat anvil
thanks for the really detailed info! We might have to investigate further downstream, could even be something weird happening between us and cloudflare
warm walrus
Something came to mind - do you all see any pattern in how much time passes betw...
it the normal time of response
warm walrus
thanks for the really detailed info! We might have to investigate further downst...
im tge oringanal poster
serene pasture
so this is a bit weird, but I also see that at the same time my Teams calls got better disabling QUIC, that this also gets better: https://quic.nginx.org/quic.html
that's a test harness from Nginx for testing QUIC... while running I see it fail every 2-3 runs, where a bunch of requests will go green (QUIC) to red (failed). Black means normal TCP based HTTP/2
this happens on multiple networks (fiber, cellular tether), OSes (Windows and macOS), and browsers (Chrome and Brave)
serene pasture
happens on Opera as well, maybe https://quic.nginx.org/quic.html isn't the best test, could be giving failures on all platforms for some reason once in a while
wheat anvil
that's a test harness from Nginx for testing QUIC... while running I see it fail...
Hmm, not sure how to interpret those results. I tried that test and it was 100% good for me on Firefox, but occasionally failed on Chrome. So that would imply something weird about Chrome and nginx, which seems hard to believe. Do Firefox users never get a network error?
serene pasture
Hmm, not sure how to interpret those results. I tried that test and it was 100% ...
Firefox always downgrades rapidly to HTTP/2 instead of failing.
I'm going to build up a sort of isolated test harness to see if I can trace this down
I think there's too many variables in place with some of these methods
serene pasture
@wheat anvil I wonder if you can detect instances of this, and tell from user agents or such what the common denominator could be
serene pasture
I'm really quite confused on this, and since everything is handled by QUIC I can't really see anything much, e.g., sequence on the packets, no dropped packets, etc
wheat anvil
I have some data on how often it's happening (which is overall not often) but nothing obvious. Firefox has a different message for network errors, but hard to tell if it's the same issue or not.
serene pasture
any hints on certain OS or browser combos?
wheat anvil
any hints on certain OS or browser combos?
Nothing stands out. Affects the top OSes at about the same rate. Reported by all Chrome-based browsers. It's possible that there's regular network errors (like getting disconnected) mixed with actual network weirdness.
serene pasture
the really odd thing is I can only seem to get this to be an issue on IPv6, @wheat anvil do you see any correllation to IPv6 vs IPv4 clients?
serene pasture
I really was suspect of my network, it's complex: mixed 10G/2.5G/1G switching, Mikrotik router with "bypassed" fiber feed (via a custom PON instead of the AT&T router), two backup internet feeds, Aruba WiFi, etc... it's not normal 🙂 So I really wanted to elimiinate that from the equation.
So I did just that... I took my laptops on macOS 14.3 and also an older Window 10 PC and did two tests:
... and it still happens when using QUIC + IPv6 in all the above configurations. If I force IPv4, it's fine even across QUIC.
Browsers:
I'm really stuck now, hoping someone else can replicate my findings and maybe this can feed a bug report of some sort. Really curious that for me MS Teams seems to behave better not on QUIC, but maybe that's because it's based on Chromium as well!
serene pasture
@warm walrus are you dual stack (IPv6)? can check at https://test-ipv6.com/
serene pasture
I have been experiencing lately while using ChatGPT. The problem primarily invol...
same to you, are you dual stack? https://test-ipv6.com/
wheat anvil
the really odd thing is I can only seem to get this to be an issue on IPv6, <@11...
Unfortunately the client-side code can't tell if it's v4 or v6 (not without some special test requests like those ipv6 test pages) so I don't know. Maybe there's a way to correlate it in the logs somehow but in any case I'll pass your info onto the team.
serene pasture
Unfortunately the client-side code can't tell if it's v4 or v6 (not without some...
Oh, I thought you were looking at request logs as well. Yeah that makes sense.
serene pasture
So confusing... I made things simpler by hosting a 1GB file on Cloudflare, and force a HTTP/3 QUIC connection via curl (found a static version with HTTP/3 complied in) and it only fails to download slowing down to a crawl on my on my home network across AT&T Fiber, but it's fine across NAT66 to my TMO cellular, off cellular hotspot, etc. What the heck is going on.
At least it's simpler to reproduce now, will tackle more in the next few days just very busy atm so I can't really dedicate much time.
What the heck, tested on another machine (same macOS + curl binary), network with Mikrotik gear and IPv6 via Comcast, and it works there.
serene pasture
All hunches from network engineers I've talked to was an issue from AT&T to Cloudflare
I've bypassed everything on my end to raw AT&T connection, and tend to agree
Today performance is muc better
h3/QUIC IPv4
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 976M 100 976M 0 0 19.0M 0 0:00:51 0:00:51 --:--:-- 19.4M
h3/QUIC IPv6
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 976M 100 976M 0 0 6714k 0 0:02:28 0:02:28 --:--:-- 5317k
h2 IPv4
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 976M 100 976M 0 0 28.7M 0 0:00:33 0:00:33 --:--:-- 23.2M
h2 IPv6
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 976M 100 976M 0 0 27.7M 0 0:00:35 0:00:35 --:--:-- 30.2M```earlier in the day h3/QUIC IPv6 was faster than IPv4, now it is slower
I'm in Georgia, the other person also saw terrible performance yesterday (500KB/s or lower) an AT&T in Florida
serene pasture
h3/QUIC IPv6 - 750.11 KB/s
h2 IPv4 - 61336.74 KB/s
h2 IPv6 - 28558.57 KB/s```feels like this is an AT&T to Cloudflare peering issue for IPv6 traffic tbh
and QUIC is just a tempermental protocol, since it can't rely on the decades of IPv4 attention and tuning, it has to reimplement all the things in new software stacks now on UDP
wheat anvil
I happen to have AT&T Fiber at home too and IPv6 will stop working every couple months until I reboot the gateway lol
Hopefully some of the networking improvements we're making will work around this issue
serene pasture
AT&T's IPv6 support is quite trash, they apparenly charge too much for bandwidth on peering and have poor peering agreements
everything I can see from my tests, mtr, etc is bad AT&T to Cloudflare (and others) peering
I hate to do it, but I disabled IPv6 for now
QUIC is also sort of junk when connections have issues it seems, they had to re-implement congestion control, retries, etc and it seems to fare worse for me than non-QUIC -- which is actually not what is expected as QUIC was supposed to be good at this stuff
hollow dew
I give up on trying to do anything I'm so done at this point this has happened s...
Which AI will you be moving to? I'm at a point now where I'm over this drama as well. I need to move away from ChatGPT. Any recommendations will be greatly appreciated.
Also, I've been getting all these same errors as everyone else on the thread. I live in Thailand - so my networks etc are different to yours - I definitely don't think the networks are to blame. This is a GPT server situation I think.
serene pasture
I've 100% verified my issue with constant network errors is due to network issues, but your situation may be different.
Try pinging the servers, run mtr, etc to see
Also try disabling QUIC in chrome://flags and search for QUIC
that is the easiest to do right now
hollow dew
Errrrm is there some kind of article I can refer to for step by step on how to do this @serene pasture
limpid python
I’m also running into this issue. It only occurs in one of my chats and not all of them (sorry for the terrible picture, I don’t have any screenshot software on my new pc yet)
serene pasture
Best to try disabling QUIC. If you are on Chrome or Brave type in "chrome://flags" into the browser address bar then search for QUIC, disable and restart. Let me know!
limpid python
same error, only in the one chat. tried disabling QUIC.
serene pasture
Also be curious what ISP you are on and what the result of test-ipv6.com and speed.cloudflare.com is especially the latter for packet drop
Make sure to restart the browser entirely too after changing flags
limpid python
10/10 on ipv6 test, running my internet through Mediacom. Also i was mistaken it's no longer a QUIC error but an http2 protocol error.
serene pasture
Ah. Hmm packet loss to Cloudflare at all?
Another step is disabling IPv6 although only would help if the IPv4 routes are better. That's my case with AT&T where IPv6 is trash especially in evenings.
To disable that you should search per your OS/router.
limpid python
gotcha, i just find it very interesting that my other chats that ive been using for months work fine but this one that i started recently (within the last) is having the issue.
serene pasture
That is a bit odd. Does it work on the same connection but on a different device such as phone?
limpid python
no that's also the interesting thing, its a shared network error on my phone (on cellular) and my pc (on hardwire connection) only for this one chat.
serene pasture
So that is weird
Have to head out. Maybe restart a new chat. Could be a bug on their end in that case. Hmm!
limpid python
I appreciate your attempts at helping. See you friend.
hollow dew
I've started a new thread since nothing else was working. This is now my 3rd thrad in 2 days that I need to retrain with tone of voice and providing it with documents of where we left off in previous threads. Irritating as hell. It's been recommended to me to move to Gemini so I'll be testing that out and hopefully moving to a more stable AI with them.
serene pasture
yeah this is frustrating for sure. I've solved my issue by disabling IPv6 at least, but hope you guys trace down your issues too
tawny ibex
Same here (network error). Plus, it seems that when using a document for evaluation, the LLM can't seem to find items marked in the document like it did 9 other times (different marked sections).
buoyant granite
I have the same trouble and Open AI support doesn't answer me
wheat anvil
no that's also the interesting thing, its a shared network error on my phone (on...
When you say it's only for that one chat, do you get "network error" 100% of the time for that chat? If so, would you be willing to share that chat with me via DM? (No problem if not, but it'd be interesting to see what's wrong with that chat)
serene pasture
Update after quite a while is that disabling IPv6 solved my issue on AT&T. QUIC is for some reason very temperamental on unreliable connections, which is very unusual, since the protocol supposed to be designed to be resilient to that. Oh well another half baked idea by Google.
sage ivy
I've been having this same issue off and on for over a month now. I disabled IPv6, and I'll monitor how things go today. Looks promising so far. Thanks @serene pasture for your detailed work here. Edit: AT&T here too. Maybe they should have stuck with telephone and telegraph. It's even in their name.
sage ivy
Update, that seemed to work for a few hours and then went back to how it was before. I fully disabled IPv6.