VPN: A/B test whether the NordVPN tunnel is degrading streams #118
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Why
Owner, 2026-11-08: no game has played cleanly for more than 5–10 minutes today. The prime suspect is the NordVPN dedicated-IP stack:
docker-compose.ymlhas been the gluetun/WireGuard stack since6bcfd2b/02bed23on develop, live since about 07:00Z today (sb-gluetun,sb-flaresolverrand the app all recreated 16:02Z with #116's deploy).This test is meant to prove or rule out the VPN before we act on it, because today has other suspects too:
hls.lshls.stsent 800 KB segments slower than their 4 s through both exits (under 2 Mbit/s), while the same tunnel did 27–45 Mbit/s to OVH.hls.lshls.stpublishing at 0.86–0.93×, measured from the Mac outside the VPN too.What to measure
A channel stutters for one of two reasons, and only the second can be the VPN's fault:
Also count challenges and solve times for the hubs and the solver, since the exit address is new to Cloudflare.
Step 0: look back at today (cheap, confounded)
Using
/api/events, compare swaps, recoveries and filler seconds per game-hour before and after ~07:00Z. This is only a hint, because the 08:48Z and 16:02Z deploys fall in the same window.Test 1: one stream over both paths at the same time (no downtime)
network_mode: container:sb-gluetun)./app/data/readiness.json; read it with Portainer's archive API.hls.lshls.st,edgestream*.pro/ the hls_co JSON on tiktokcdn, and any others live).Test 2: the whole stack with and without the VPN
docker-compose.ymlfrom develop, so each switch is a commit that swaps indocker-compose.novpn.yml(asfd1d5a6did), followed by a redeploy.Grade.slowgradesseconds_since_new_segmentgaps over 8 s/api/events, the container log'sclock: seq N splices/recovered onlines, and/api/admin.sb-gluetunbefore bringing up the plain stack, because it holds 7000 and 8191.Done when
If the tunnel is the cause, check next
WIREGUARD_MTUin gluetun) and fragmentation.Verdict (2026-11-08, 17:50–18:55Z)
The tunnel doesn't slow streams. The VPN stack hurts in two other ways:
BLOCK_MALICIOUS, on by default) answers REFUSED for the hosts of most players and of two of the fastest stream CDNs.streame.center,embedcenter.st,edgestream3.pro,hls.meth-streams-hd.cfd,s1.saltext.ch,play.aurasportsplayer.cfd,frame.embedme.cfd, and ad/analytics hosts.manifest: upstream unreachable: ConnectErrorlink failures come from (iohit.st, shd247.world): the stream host never resolves.s1.saltext.ch's) are tied to the address that opened the player.The result:
hls.lshls.st, which is slow at peak on any path.edgestream3.proandhls.meth-streams-hd.cfd, which deliver a segment in 4–10% of its play time.How it was measured
Two short-lived probe containers on the VM ran the app's own image, fetch code and browser code. One shared
sb-gluetun's network (tunnel, exit Ashburn VA); the other was on Docker's bridge (direct, the home address). They tested the same targets at the same moment, with no downtime, and were removed afterwards.The tunnel itself is healthy
Through the tunnel a TCP connect answers in about 14 ms whatever the destination, as if NordVPN's server accepts connections itself and relays them. The TLS handshake is the real end-to-end figure, and it is close on both paths.
hls.lshls.st, 20 minutes at the 1 PM kickoff (18:02–18:22Z)
Every newly listed segment was fetched on each path at the same moment.
The same server, the same slowness. The tunnel adds about 3% at the median and a heavier tail. (The Saints link expired, with 403s on both paths at once.)
Player pages, 33 sites, read once each from both paths at the same moment
This used the app's own browser code, before any solver handoff.
With the refused names looked up at 1.1.1.1 inside the tunnel:
The owner's split: find at home, play through the VPN
Sixteen streams were found from home, then fetched for 2–3 minutes. The tunnel side had its DNS fixed the same way.
Step 0: today's history (confounded by 1.2.0, #116, #119 and the game mix)
Options
BLOCK_MALICIOUS=offon gluetun, redeployRecommendation:
Test 2 (the whole-stack A/B) wasn't run: the same-moment tests above show the cause directly. After A, one game window will confirm it, measured by the longest clean stretch, swaps, recoveries and filler per game-hour, and the check success rate.
After turning gluetun's DNS filter off (PR #120, merge
7700913, deployed 19:02:44Z)The owner chose option A. The probes were then repeated, and the app was watched on its own.
The same page test, 33 player sites, both paths at the same moment
The tunnel now matches direct site for site.
Correction to my earlier report: the 7 edgestream3.pro sites do give streams through the VPN. "edgestream3's player refuses the NordVPN address" was wrong: it came from my test's DNS workaround, not from the site. So option A covers everything that works from home, and options B, B+ and C are moot for coverage.
Playback through the tunnel of streams found through the tunnel (3 minutes, 19:11Z)
The app's own numbers
What the tunnel still costs: connection setup under bursts
upstream._client()per request). That load includes the watch's polls of up to 24 links.Verdict
Partly the VPN stack, not the tunnel.
edgestream3.pro,hls.meth-streams-hd.cfdands1.saltext.chstreams need, and that left the app onhls.lshls.st, which is slow at peak on any path.The stack stays on the VPN with the filter off; the reason is in the compose notes (#120). Closing, as the fix is merged.