fix: Using sliding window rate limiter

sdk/python/feast/infra/online_stores/contrib/cassandra_online_store/cassandra_online_store.py

@@ -242,6 +242,24 @@ def __exit__(self, exc_type, exc_val, exc_tb):
         pass
 
 
+class SlidingWindowRateLimiter:
+    def __init__(self, max_calls, period):
+        self.max_calls = max_calls
+        self.period = period
+        self.timestamps = [0] * max_calls
+        self.index = 0
+
+    def acquire(self):
+        now = time.time()
+        window_start = now - self.period
+
+        if self.timestamps[self.index] <= window_start:
+            self.timestamps[self.index] = now
+            self.index = (self.index + 1) % self.max_calls
+            return True
+        return False
+
+
 class CassandraOnlineStore(OnlineStore):
     """
     Cassandra/Astra DB online store implementation for Feast.
@@ -410,7 +428,8 @@ def on_failure(exc, concurrent_queue):
             raise Exception("Error writing a batch") from exc
 
         write_concurrency = config.online_store.write_concurrency
-        concurrent_queue: queue.Queue = queue.Queue(maxsize=write_concurrency)
+        rate_limiter = SlidingWindowRateLimiter(write_concurrency, 1)
+        concurrent_queue: queue.Queue = queue.Queue(maxsize=write_concurrency / 5)
 
         project = config.project
         ttl = (
@@ -444,6 +463,12 @@ def on_failure(exc, concurrent_queue):
                     timestamp,
                 )
                 batch.add(insert_cql, params)
+
+            # Wait until the rate limiter allows
+            if not rate_limiter.acquire():
+                while not rate_limiter.acquire():
+                    time.sleep(0.001)
+
             future = session.execute_async(batch)
             concurrent_queue.put(future)
             future.add_callbacks(
@@ -460,6 +485,7 @@ def on_failure(exc, concurrent_queue):
             # this happens N-1 times, will be corrected outside:
             if progress:
                 progress(1)
+        # Wait for all tasks to complete
         while not concurrent_queue.empty():
             time.sleep(0.001)