Fix Fetcher.PartitionRecords to handle fetch_offset in the middle of compressed messageset (#1239)

2 files changed, 31 insertions, 3 deletions

diff --git a/kafka/consumer/fetcher.py b/kafka/consumer/fetcher.py
index b86c8ec..f552038 100644
--- a/kafka/consumer/fetcher.py
+++ b/kafka/consumer/fetcher.py
@@ -923,12 +923,17 @@ class Fetcher(six.Iterator):
             self._sensors.fetch_throttle_time_sensor.record(response.throttle_time_ms)
         self._sensors.fetch_latency.record((recv_time - send_time) * 1000)
 
-    class PartitionRecords(six.Iterator):
+    class PartitionRecords(object):
         def __init__(self, fetch_offset, tp, messages):
             self.fetch_offset = fetch_offset
             self.topic_partition = tp
             self.messages = messages
-            self.message_idx = 0
+            # When fetching an offset that is in the middle of a
+            # compressed batch, we will get all messages in the batch.
+            # But we want to start 'take' at the fetch_offset
+            for i, msg in enumerate(messages):
+                if msg.offset == fetch_offset:
+                    self.message_idx = i
 
         def discard(self):
             self.messages = None
diff --git a/test/test_fetcher.py b/test/test_fetcher.py
index 64eec1b..86d154f 100644
--- a/test/test_fetcher.py
+++ b/test/test_fetcher.py
@@ -7,7 +7,7 @@ import itertools
 from collections import OrderedDict
 
 from kafka.client_async import KafkaClient
-from kafka.consumer.fetcher import Fetcher, NoOffsetForPartitionError
+from kafka.consumer.fetcher import ConsumerRecord, Fetcher, NoOffsetForPartitionError
 from kafka.consumer.subscription_state import SubscriptionState
 from kafka.metrics import Metrics
 from kafka.protocol.fetch import FetchRequest
@@ -282,3 +282,26 @@ def test__handle_offset_response(fetcher, mocker):
     fetcher._handle_offset_response(fut, res)
     assert fut.failed()
     assert isinstance(fut.exception, NotLeaderForPartitionError)
+
+
+def test_partition_records_offset():
+    """Test that compressed messagesets are handle correctly
+    when fetch offset is in the middle of the message list
+    """
+    batch_start = 120
+    batch_end = 130
+    fetch_offset = 123
+    tp = TopicPartition('foo', 0)
+    messages = [ConsumerRecord(tp.topic, tp.partition, i,
+                               None, None, 'key', 'value', 'checksum', 0, 0)
+                for i in range(batch_start, batch_end)]
+    records = Fetcher.PartitionRecords(fetch_offset, None, messages)
+    assert records.has_more()
+    msgs = records.take(1)
+    assert msgs[0].offset == 123
+    assert records.fetch_offset == 124
+    msgs = records.take(2)
+    assert len(msgs) == 2
+    assert records.has_more()
+    records.discard()
+    assert not records.has_more()