v2.3.3

[2.3.3] — 2026-10-02

Fixed

• The first call after a few idle minutes no longer fails with "no response … connection
  failed". A NAT gateway or proxy between the client and Sisense drops idle keep-alive
  connections without closing them; the pooled connection then dies with RemoteDisconnected
  before the server saw the request. Two changes: every pooled socket now has TCP keepalive
  (probe after 60 s idle, then every 30 s), which keeps the mapping alive so connections do not
  go stale; and the retry policy allows one retry on connection errors and one on read errors
  (connect=1, read=1, previously both 0). urllib3 applies read retries to idempotent methods
  only, so a POST is never sent twice; POST is protected by the keepalive. Seen on an EC2 host
  without a public IP behind an AWS NAT gateway (350 s idle timeout).

Downstream generators: no introspection-surface changes.