rgw/multisite: fix uninitialized LatencyMonitor average and use exponentially weighte...

LatencyMonitor::total was declared without an initializer. Since
std::chrono::duration's default constructor leaves the value indeterminate,
the very first add_latency() call adds a real sample to garbage, producing a
huge average that immediately triggers the "OSD cluster is overloaded" warning
within seconds of RGW startup, before any actual slow ops occur.

Additionally, the old implementation uses a naive lifetime average
(total/count) that could slow the recovery from a transient slow-ops
episode. Once poisoned, the average stayed high for a long time,
keeping the throttling sync concurrency to 1.

So, also replace the naive lifetime average in LatencyMonitor with an
exponentially weighted moving average (alpha=0.15). With the weighted average,
after a series of  normal lock operations a past spike's influence decays faster,
allowing concurrency to recover without an RGW restart.

Fixes: https://tracker.ceph.com/issues/76308
Signed-off-by: Oguzhan Ozmen <oozmen@bloomberg.net>

diff --git a/src/rgw/driver/rados/rgw_cr_rados.h b/src/rgw/driver/rados/rgw_cr_rados.h
index 1aa9f17e335faa2914ce211db90ffe7e9ed113d8..481a7283e5ad17b660f5412efcc786b23d48ad7e 100644 (file)
--- a/src/rgw/driver/rados/rgw_cr_rados.h
+++ b/src/rgw/driver/rados/rgw_cr_rados.h
@@ -1523,20 +1523,30 @@ public:
 /// \warning This class is not thread safe. We do not use a mutex
 /// because all coroutines spawned by RGWDataSyncCR share a single thread.
 class LatencyMonitor {
-  ceph::timespan total;
-  std::uint64_t count = 0;
+  ceph::timespan avg{ceph::timespan::zero()};
+  bool initialized = false;
+  // Weight for new samples in running average. Recent samples matter
+  // most; after ~20 new samples a past spike decays to <4%.
+  // Example: if avg is poisoned at 30s but real latency is 0.1s,
+  // after 20 good samples the avg drops to ~1.2s (fully recovered).
+  static constexpr double alpha = 0.15;
 
 public:
 
   LatencyMonitor() = default;
   void add_latency(ceph::timespan latency) {
-    total += latency;
-    ++count;
+    if (!initialized) {
+      avg = latency;
+      initialized = true;
+    } else {
+      avg = ceph::timespan(
+        static_cast<ceph::timespan::rep>(
+          alpha * latency.count() + (1.0 - alpha) * avg.count()));
+    }
   }
 
   ceph::timespan avg_latency() {
-    using namespace std::literals;
-    return count == 0 ? 0s : total / count;
+    return avg;
   }
 };