OVH Mumbai (YNM1): one border path to Tata loses all traffic for ~4 min every ~20 min

Since 06:00 UTC today (Sat 31 Oct), our dedicated server in OVH Mumbai (YNM1) has lost all traffic on one of OVH’s two paths to Tata Communications for 3–4 minutes roughly every 20 minutes, with more than 15 outages recorded. We traced the issue to the route via bom-ynm1-sbb1-8k (103.5.12.8) → 10.200.4.7; the alternative path via bom-mb2-sbb1-8k remains unaffected. A ticket is open with complete evidence, but India phone support operates only on weekdays and no one has drained the path. Has anyone else on OVH Mumbai observed this?

What we see

An outage occurs every 18–23 minutes, lasting 3–4 minutes and lengthening through the day.
During each event, traffic to any destination behind Tata stops: our telecom carrier (via Tata AS4755) and one server in Germany (via Tata AS6453). Google DNS and our US server show 0% loss throughout.
Only the outgoing direction fails. Packets from the carrier continue arriving; ours do not reach them.
The server itself appears healthy: no interface errors, socket drops, or kernel log entries.

How we isolated it

Fixed-flow traceroutes (UDP from 10 fixed source ports so each flow keeps one ECMP path) run through an outage:

Path normal during outage
A: 103.5.12.10 bom-mb2-sbb1-8k -> 10.200.4.5 -> Tata ~0% 0 of 76 probes lost
B: 103.5.12.8 bom-ynm1-sbb1-8k -> 10.200.4.7 -> Tata ~0% 49 of 49 probes lost

Flows mapped to path B fail just after 10.200.4.7 for the entire outage; flows on path A remain unaffected. ICMP mtr may appear entirely clean during an outage if its flow happens to map onto path A, which likely explains why standard tests overlook this.

Our guess

A link on path B toward Tata fails in one direction and stays active until a routing session times out (about 3 minutes), then returns and fails again around 20 minutes later. The 20-minute cycle resembles a default ARP or IS-IS timer, though that remains speculative.

The ask- OVH team: please route via bom-ynm1-sbb1-8k / 10.200.4.7 toward Tata until repairs are complete. Ticket #[number] includes attached ip/ethtool output, mtr readings before, during, and after, plus per-flow traceroutes.
Others in Mumbai: do you observe the same 20-minute pattern toward Tata, Airtel, or Jio?

This is the second incident in 10 days; on 21 Oct the same exit suffered a 34-minute total outage. OVH’s status page reports nothing for either event.