Free (AS12322, France) routed to lhr: about half of TCP connections to Fly anycast IPs fail

Since at least 2026-10-20 15:30 UTC, about half of the TCP connections from a Free (AS12322) fiber line in France to Fly’s anycast addresses never complete. fly-request-id shows that these requests land on lhr. Deploys break because machine updates through the Machines API fail with MachineMutationUncertain and MachineLeaseLost (HttpClientError and TimeoutError on lease acquisition and refresh).

Measurements from the Free line: curl -4, 3 s connect timeout, up to 20 attempts per endpoint, around 16:00–16:20 UTC.

Endpoint Successes
api.machines.dev (137.66.34.115) 9/20, then 14/20
fly.io (37.16.18.81) 9/20
registry.fly.io (77.83.143.221) 4/6
api.github.com, same host 20/20

The failed attempts never get a TCP connection. DNS is not involved: the Freebox resolver, 1.1.1.1 and 8.8.8.8 all return these same addresses.

From another French ISP at the same time, the same addresses route to cdg, with 20/20 successes on every endpoint.

Is anyone else on Free seeing this? Could someone at Fly check the path between AS12322 and lhr, or why Free traffic lands on lhr instead of cdg?

Hi… Usually you have to check this yourself; at least, they nearly always ask for the following in situations like this:

(That’s one of many, many examples in the forum.)

[Be sure to also include the response headers from curl in your forum post, incidentally, :sweat_smile:.]

Could you include a traceroute / mtr report from your local machine to our IPs? Based on RIPE probes it doesn’t look like the route is problematic and it seems like AS12322 is going to cdg as expected:

Traceroute from 82.64.77.142 to 37.16.18.81
Using RIPE, Probe 1009235

STARTED QUERY AT 2026/10/20 16:53:52 UTC

Fetching Measurement: 225366026
Traceroute from 192.168.1.138 to 37.16.18.81 (37.16.18.81):
1 * 192.168.1.254 0.251ms *  
2 194.149.162.224 2.128ms 1.565ms 2.161ms  
3 * * *  
4 prs-b3-link.ip.twelve99.net (62.115.46.68) 11.474ms 7.633ms 7.336ms  
5 datacamp-ic-382929.ip.twelve99-cust.net (213.248.75.173) 7.529ms 7.718ms 7.98ms  
6 unn-185-156-45-114.cdn77.com (185.156.45.114) 7.882ms 7.743ms 8.176ms  
7 vl221.par-itx5-dist-2.cdn77.com (79.127.195.36) 7.336ms 7.967ms 7.377ms  
8 ip-37-16-18-81.customer.flyio.net (37.16.18.81) 7.857ms 7.93ms 7.73ms  

Completed in 26.52s

Thanks @PeterCxy, here’s the mtr from the affected machine, run while connections were failing (2026-10-20 16:44 to 16:50 UTC). Unlike your RIPE probe, this line leaves Free through AS3356 in London rather than Telia in Paris, so it lands on lhr.

mtr, ICMP:

HOST: nuc                                                          Loss%   Snt Drop   Last   Avg  Best  Wrst StDev
  1. AS???    _gateway (192.168.5.1)                                5.0%    20    1    1.8   1.8   1.6   2.9   0.3
  2. AS???    192.168.1.254                                        20.0%    20    4    2.0   2.0   1.6   3.4   0.4
  3. AS???    station15.multimania.isdnet.net (194.149.174.112)    85.0%    20   17    4.6   3.8   3.2   4.6   0.7
  4. AS???    station17.multimania.isdnet.net (194.149.174.114)    80.0%    20   16    3.6   4.3   3.5   5.7   1.0
  5. AS12322  212.27.35.0                                          40.0%    20    8    4.2   4.2   3.7   4.6   0.3
  6. AS???    ???                                                  100.0    20   20    0.0   0.0   0.0   0.0   0.0
  7. AS???    ???                                                  100.0    20   20    0.0   0.0   0.0   0.0   0.0
  8. AS3356   ae1.3115.edge7.lon1.neo.colt.net (171.75.9.213)      70.0%    20   14   34.3  15.4  11.0  34.3   9.3
  9. AS3356   217.163.122.26                                        0.0%    20    0   11.3  11.6  10.9  12.6   0.4
 10. AS???    vl251.lon-eq7-dist-2.cdn77.com (79.127.195.180)      55.0%    20   11   10.7  11.5  10.7  11.9   0.4
 11. AS40509  ip-137-66-34-115.customer.flyio.net (137.66.34.115)   0.0%    20    0   11.6  11.4  10.9  12.8   0.4

mtr, TCP 443 (-T -P 443):

HOST: nuc                                                          Loss%   Snt Drop   Last   Avg  Best  Wrst StDev
  1. AS???    _gateway (192.168.5.1)                                0.0%    20    0    1.6   2.0   1.5   7.9   1.4
  2. AS???    192.168.1.254                                         5.0%    20    1    1.9   2.2   1.9   3.8   0.5
  3. AS???    station15.multimania.isdnet.net (194.149.174.112)    15.0%    20    3  1009. 1631.   3.5 5143. 1777.3
  4. AS???    station17.multimania.isdnet.net (194.149.174.114)    25.0%    20    5  5121. 2130.   3.6 5158. 2068.1
  5. AS12322  212.27.35.0                                           0.0%    20    0  1061. 569.5   3.7 3079. 1022.0
  6. AS???    ???                                                  100.0    20   20    0.0   0.0   0.0   0.0   0.0
  7. AS???    ???                                                  100.0    20   20    0.0   0.0   0.0   0.0   0.0
  8. AS3356   ae1.3115.edge7.lon1.neo.colt.net (171.75.9.213)      15.0%    20    3  2603. 647.2  10.4 5619. 1537.9
  9. AS3356   217.163.122.26                                        5.0%    20    1  5631. 1212.  11.2 5631. 1900.3
 10. AS???    vl251.lon-eq7-dist-2.cdn77.com (79.127.195.180)      15.0%    20    3  5702. 1337.  11.0 5702. 1899.4
        vl251.lon-eq7-dist-1.cdn77.com (79.127.195.179)
 11. AS40509  ip-137-66-34-115.customer.flyio.net (137.66.34.115)   5.0%    20    1  6652. 1256.  11.2 6682. 2209.8

ICMP to 137.66.34.115 lost nothing. TCP 443 to it averaged 1.26 s and peaked at 6.7 s, because initial SYNs went unanswered and had to be retransmitted. Same machine, same time: the TCP mtr to api.github.com (through DE-CIX Frankfurt) averaged 13.3 ms, and 59 of 90 connects with a 0.9 s timeout failed to 137.66.34.115, against 0 of 90 to 140.82.121.5.

curl with flyio-debug, after three 5 s connect timeouts:

< server: Fly/42494bfe9c (2026-10-15)
< fly-request-id: 01M5D35HVM7YPPN348ETQM996Y-lhr
< flyio-debug: {"n":"edge-dp-lhr1-7117","nr":"lhr","ra":"<redacted>","rf":"Verbatim","sr":"lhr","sdc":"lon1","sid":"830d1ef74e64d8","st":0,"nrtt":1,"bn":"worker-cf-lon1-4e23","mhn":null,"mrtt":null}

Update at 2026-10-21 10:38 UTC: same London path, still lhr, but 0 of 90 connects fail now, so it comes and goes. I can DM you the source IP if you want to check the lhr edge for that window, or find a RIPE probe closer to this line.

Hi! Just to provide an update, it looks like your ISP’s routing policy prefers Level 3 in some parts of the country while not so in other parts. This is slightly difficult for us to tune purely from our side so that it goes exactly to Paris (CDG). The connection and packet loss issues you have seen is probably related to Level 3 as well.

We’ll work with our upstreams to see what we can do about this.

Thanks for digging into it! Connections have been clean since the 21st on our side, I’ll post here if it comes back.