prestodb/presto · error · PrestoException

KAFKA_SPLIT_ERROR

KAFKA_SPLIT_ERROR

Error message

Cannot read data from topic '%s', partition '%s', startOffset %s, endOffset %s, leader %s 

What it means

openFetchRequest wraps any exception from the Kafka consumer's fetch/send call (other than PrestoException) into KAFKA_SPLIT_ERROR, reporting the topic, partition, offset range and leader. This means Presto could not read data from the Kafka broker for this split — connectivity, broker availability, offset-range, or serialization issues inside the raw consumer API.

Source

Thrown at presto-kafka/src/main/java/com/facebook/presto/kafka/KafkaRecordSet.java:328

                if (messageAndOffsetIterator == null) {
                    String threadName = Thread.currentThread().getName();

                    if (consumer == null) {
                        consumer = consumerManager.createConsumer(threadName, split.getLeader());
                    }

                    TopicPartition topicPartition = new TopicPartition(split.getTopicName(), split.getPartitionId());
                    consumer.assign(ImmutableList.of(topicPartition));
                    consumer.seek(topicPartition, cursorOffset);
                    ConsumerRecords<ByteBuffer, ByteBuffer> records = consumer.poll(POLL_TIMEOUT);
                    messageAndOffsetIterator = records.records(topicPartition).iterator();
                }
            }
            catch (Exception e) { // Catch all exceptions because Kafka library is written in scala and checked exceptions are not declared in method signature.
                if (e instanceof PrestoException) {
                    throw e;
                }
                throw new PrestoException(
                        KAFKA_SPLIT_ERROR,
                        format(
                                "Cannot read data from topic '%s', partition '%s', startOffset %s, endOffset %s, leader %s ",
                                split.getTopicName(),
                                split.getPartitionId(),
                                split.getStart(),
                                split.getEnd(),
                                split.getLeader()),
                        e);
            }
        }
    }
}

View on GitHub (pinned to 55bb57d202)

Solutions

1. Verify brokers are reachable from Presto worker nodes (telnet broker:9092) and kafka.nodes config is correct
2. Check the topic/partition still exists and offsets are within the retained log (kafka-run-class kafka.tools.GetOffsetShell)
3. Fix advertised.listeners on the brokers so returned host:port are resolvable by workers
4. Retry the query if it was a transient leader election

Example fix

// before
kafka.nodes=internal-broker:9092
// after - resolvable from workers, multiple nodes
kafka.nodes=broker1.internal:9092,broker2.internal:9092

// preflight: check topic reachability from worker
kafka-topics.sh --bootstrap-server broker:9092 --describe --topic <topic>

null

try { readSplit(); } catch (PrestoException e) {
  if (e.getErrorCode() == KAFKA_SPLIT_ERROR.toErrorCode()) {
    // verify broker connectivity and offset range, then retry query
  }
}

When it happens

Trigger: KafkaRecordSet.advanceNextPosition -> openFetchRequest calls consumer.send/poll on a KafkaConsumer; the Kafka scala client throws (broker unreachable, unknown topic/partition, offset out of range, leader changed mid-scan).

Common situations: Broker down or firewall blocking broker port from workers; topic deleted/renamed between split generation and read; advertised.listeners misconfiguration so workers resolve broker hostnames they cannot reach; endOffset beyond log retention (records deleted).

Related errors

• GENERIC_INTERNAL_ERROR
• KAFKA_SPLIT_ERROR
• unexpected internal column '%s'
• KAFKA_SCHEMA_ERROR
• Invalid Kafka Offset start/end pair: %s - %s

AI-assisted analysis of prestodb/presto@55bb57d202 (2026-09-04). Data as JSON: /api/errors/ac3be5c818f020d7. Report an issue: GitHub.