From 6d71a4fd3158e37f21dde31eeff69866f404cc03 Mon Sep 17 00:00:00 2001
From: Arseny Sher <sher-ars@yandex.ru>
Date: Thu, 1 Feb 2024 16:58:14 +0300
Subject: [PATCH] Ignore zero sized cluster size in pageserver feedback.

To take into account that non zero shards send 0 size. Generally we should make
reporting sharding aware, but for now this will avoid breaking size limit after
https://github.com/neondatabase/neon/pull/6567
---
 pgxn/neon/walproposer_pg.c | 16 +++++++++++++++-
 safekeeper/src/send_wal.rs | 10 ++++++----
 2 files changed, 21 insertions(+), 5 deletions(-)

diff --git a/pgxn/neon/walproposer_pg.c b/pgxn/neon/walproposer_pg.c
index 61a2a54809..ad463ebcf8 100644
--- a/pgxn/neon/walproposer_pg.c
+++ b/pgxn/neon/walproposer_pg.c
@@ -1804,11 +1804,21 @@ walprop_pg_finish_sync_safekeepers(WalProposer *wp, XLogRecPtr lsn)
 static void
 GetLatestNeonFeedback(PageserverFeedback *rf, WalProposer *wp)
 {
-	int			latest_safekeeper = 0;
+	int			latest_safekeeper = -1;
 	XLogRecPtr	last_received_lsn = InvalidXLogRecPtr;
 
 	for (int i = 0; i < wp->n_safekeepers; i++)
 	{
+		/*
+		 * Non-zero shards don't known the timeline size and send zero.
+
+		 * TODO: right now we ignore all feedback from non zero shards. We
+		 * should make reporting sharding aware instead and do per shard
+		 * aggregation as any lagging shard should trigger backpressure.
+		 */
+		if (wp->safekeeper[i].appendResponse.rf.currentClusterSize == 0)
+			continue;
+
 		if (wp->safekeeper[i].appendResponse.rf.last_received_lsn > last_received_lsn)
 		{
 			latest_safekeeper = i;
@@ -1816,6 +1826,10 @@ GetLatestNeonFeedback(PageserverFeedback *rf, WalProposer *wp)
 		}
 	}
 
+	/* no feedback yet */
+	if (latest_safekeeper == -1)
+		return;
+
 	rf->currentClusterSize = wp->safekeeper[latest_safekeeper].appendResponse.rf.currentClusterSize;
 	rf->last_received_lsn = wp->safekeeper[latest_safekeeper].appendResponse.rf.last_received_lsn;
 	rf->disk_consistent_lsn = wp->safekeeper[latest_safekeeper].appendResponse.rf.disk_consistent_lsn;
diff --git a/safekeeper/src/send_wal.rs b/safekeeper/src/send_wal.rs
index ee3e4c8ead..d7ad678ca9 100644
--- a/safekeeper/src/send_wal.rs
+++ b/safekeeper/src/send_wal.rs
@@ -265,9 +265,9 @@ impl WalSendersShared {
     /// Update aggregated pageserver feedback. LSNs (last_received,
     /// disk_consistent, remote_consistent) and reply timestamp are just
     /// maximized; timeline_size if taken from feedback with highest
-    /// last_received lsn. This is generally reasonable, but we might want to
-    /// implement other policies once multiple pageservers start to be actively
-    /// used.
+    /// last_received lsn except when it is 0 (sent from non zero shards). This
+    /// is generally reasonable, but we might want to implement other policies
+    /// once multiple pageservers start to be actively used.
     fn update_ps_feedback(&mut self) {
         let init = PageserverFeedback::empty();
         let acc =
@@ -276,7 +276,9 @@ impl WalSendersShared {
                 .flatten()
                 .fold(init, |mut acc, ws_state| match ws_state.feedback {
                     ReplicationFeedback::Pageserver(feedback) => {
-                        if feedback.last_received_lsn > acc.last_received_lsn {
+                        if feedback.current_timeline_size != 0
+                            && feedback.last_received_lsn > acc.last_received_lsn
+                        {
                             acc.current_timeline_size = feedback.current_timeline_size;
                         }
                         acc.last_received_lsn =