idempotency/tests: add test for producer recovery

tests/rptest/transactions/idempotency_test.py

@@ -0,0 +1,146 @@
+# Copyright 2020 Redpanda Data, Inc.
+#
+# Use of this software is governed by the Business Source License
+# included in the file licenses/BSL.md
+#
+# As of the Change Date specified in that file, in accordance with
+# the Business Source License, use of this software will be governed
+# by the Apache License, Version 2.0
+from collections import defaultdict
+from rptest.clients.default import DefaultClient
+from rptest.clients.rpk import RpkTool
+from rptest.tests.redpanda_test import RedpandaTest
+from rptest.clients.types import TopicSpec
+from rptest.services.cluster import cluster
+from rptest.transactions.verifiers.idempotency_load_generator import PausableIdempotentProducer
+from time import sleep
+from rptest.util import wait_until
+import confluent_kafka as ck
+
+
+class IdempotentProducerRecoveryTest(RedpandaTest):
+    """This test ensures that various client implementations can recover
+    from scenario in which the producers get evicted on the brokers.
+    When a producer is evicted it loses all old state including the
+    sequence numbers. So the client attempting to produce after that
+    should still recover and be able to make progress. There are subtle
+    variations among client implementations around how they deal
+    with this situation, so this test acts a regression test ensuring that we
+    do not break this behavior."""
+    def __init__(self, test_context):
+        super(IdempotentProducerRecoveryTest,
+              self).__init__(test_context=test_context, num_brokers=1)
+        self.test_topic = TopicSpec(name="test",
+                                    partition_count=1,
+                                    replication_factor=1)
+
+    def wait_for_eviction(self, active_producers_remaining,
+                          expected_to_be_evicted):
+        def do_wait():
+            samples = [
+                "idempotency_pid_cache_size",
+                "producer_state_manager_evicted_producers"
+            ]
+            brokers = self.redpanda.started_nodes()
+            metrics = self.redpanda.metrics_samples(samples, brokers)
+            producers_per_node = defaultdict(int)
+            evicted_per_node = defaultdict(int)
+            for pattern, metric in metrics.items():
+                for m in metric.samples:
+                    id = self.redpanda.node_id(m.node)
+                    if pattern == "idempotency_pid_cache_size":
+                        producers_per_node[id] += int(m.value)
+                    elif pattern == "producer_state_manager_evicted_producers":
+                        evicted_per_node[id] += int(m.value)
+
+            self.redpanda.logger.debug(
+                f"active producers: {producers_per_node}")
+            self.redpanda.logger.debug(
+                f"evicted producers: {evicted_per_node}")
+
+            remaining_match = all([
+                num == active_producers_remaining
+                for num in producers_per_node.values()
+            ])
+
+            evicted_match = all([
+                val == expected_to_be_evicted
+                for val in evicted_per_node.values()
+            ])
+
+            return len(producers_per_node) == len(
+                brokers) and remaining_match and evicted_match
+
+        wait_until(do_wait,
+                   timeout_sec=20,
+                   backoff_sec=1,
+                   err_msg=f"Not all producers were evicted in 20secs.",
+                   retry_on_exc=False)
+
+    @cluster(num_nodes=2)
+    def test_java_client_recovery_on_producer_eviction(self):
+        rpk = RpkTool(self.redpanda)
+        rpk.create_topic(self.test_topic.name, self.test_topic.partition_count,
+                         self.test_topic.replication_factor)
+
+        workload_svc = PausableIdempotentProducer(self.test_context,
+                                                  self.redpanda)
+        workload_svc.start()
+
+        workload_svc.start_producer(self.test_topic.name,
+                                    self.test_topic.partition_count)
+        # Generate some load
+        workload_svc.ensure_progress(expected=1000, timeout_sec=20)
+
+        # Pause the producer to evict the producer sessions
+        workload_svc.pause_producer()
+
+        progress_so_far = workload_svc.get_progress().json()["num_produced"]
+
+        # Evict all producers
+        rpk.cluster_config_set("transactional_id_expiration_ms", 0)
+        self.wait_for_eviction(active_producers_remaining=0,
+                               expected_to_be_evicted=1)
+        rpk.cluster_config_set("transactional_id_expiration_ms", 3600000)
+
+        # Resume the idempotency session again.
+        workload_svc.start_producer(self.test_topic.name,
+                                    self.test_topic.partition_count)
+
+        workload_svc.ensure_progress(expected=progress_so_far + 1000,
+                                     timeout_sec=20)
+        # Verify the producer can make progress without exceptions
+        workload_svc.stop_producer()
+
+    @cluster(num_nodes=1)
+    def test_librdkafka_recovery_on_producer_eviction(self):
+        rpk = RpkTool(self.redpanda)
+        rpk.create_topic(self.test_topic.name, self.test_topic.partition_count,
+                         self.test_topic.replication_factor)
+
+        producer = ck.Producer({
+            'bootstrap.servers': self.redpanda.brokers(),
+            'enable.idempotence': True,
+        })
+
+        def produce_some():
+            def on_delivery(err, _):
+                assert err is None, err
+
+            for i in range(1000):
+                producer.produce(self.test_topic.name,
+                                 str(i),
+                                 str(i),
+                                 on_delivery=on_delivery)
+            producer.flush()
+
+        produce_some()
+
+        # Evict all producers
+        rpk.cluster_config_set("transactional_id_expiration_ms", 0)
+        self.wait_for_eviction(active_producers_remaining=0,
+                               expected_to_be_evicted=1)
+        rpk.cluster_config_set("transactional_id_expiration_ms", 3600000)
+
+        # Ensure producer can recover.
+        produce_some()