議論： Handling bounces and complaints: DSNs, enhanced status codes and feedback loops

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-16

翻訳がないため、原文を表示しています。 原文

Two sources the pipeline must join in. Many permanent failures never produce a DSN: the receiving server rejects the recipient synchronously with a 5yz reply at `RCPT TO` or after `DATA`, and it is the sending worker, not the bounce mailbox, that sees the enhanced status code; a suppression pipeline that only parses incoming DSNs misses these unless the sender's own SMTP replies are fed into the same table. On feedback loops, Gmail does not send per-message ARF reports; it offers an aggregated Feedback Loop through Postmaster Tools that keys on a `Feedback-ID` header the sender adds, and reports spam rates per identifier rather than per recipient, so step 5's 'suppress on the first abuse report' cannot be executed for Gmail addresses and the per-stream rate in step 6 is the only signal there. Microsoft's JMRP and Yahoo's CFL do send ARF, so the parser is still needed.

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-16

翻訳がないため、原文を表示しています。 原文

Step 4's rule for X.7.X, 'a signal about the sender, not the address', is right about the diagnosis and wrong about the consequence. A suppression table keyed by address has no place to record that a whole receiving domain is refusing the sender (5.7.1 from a blocklisting, 4.7.0 throttling, 4.7.1 greylisting), so the pipeline keeps sending to every address at that domain on every campaign, and each rejected attempt is more evidence for the receiver's filter; the damage is done per domain, the bookkeeping is per address, and the mismatch is what turns a temporary block into a lasting one. The unit for policy codes is the receiving domain (or its MX): pause sending to the domain, alert a human, and resume after a probe succeeds. That also fixes the observation in step 6, since the 'rate per stream' hides a domain that fails completely while the total looks fine. The rule should therefore read: address-level suppression for X.1.X and X.2.X, domain-level pause for X.7.X and X.4.X.