Improving Kafka at-least-once performance at Uber

Improving Kafka At-Least-Once
Performance
Ying Zheng, Streaming Data Team
Uber

Kafka at Uber
● Use cases
○ General Pub-Sub
○ Stream Processing
○ Ingestion
○ Database Changelog Transport
○ Logging
● Scale
○ Trillion+ messages / day
○ Tens of thousands of topics

Kafka at-least-once delivery at Uber
● Started using at-least-once delivery in 2016
● More and more services are using Kafka to pass business critical messages
○ E.g., payment events, driver profile management, and trip updates etc.
● Scope:
○ Hundreds of topics
○ Tens of thousands of messages per second

The performance issue
● Producer latency is important for at-least-once use cases
● Some use cases also require low end-to-end latency
● A simple performance benchmark shows that Kafka at-least-once performance
doesn’t scale well:
# partitions per node Throughput per node
(MB/s)
P50 latency (ms) P99 latency (ms)
4 105 4 27
40 64 7 27
800 64 18 58

The first at-least-once cluster at Uber
● For a dedicated at-least-once cluster, the per-node traffic was ~ 30x lower than
the other Kafka clusters
● Users were manually approved
● There are some very large use cases (hundreds thousands of messages per
second) waiting. But, the cluster could not handle that much traffic.

The performance improvement project
● Make sure at-least-once Kafka cluster can handle as much traffic as the regular
clusters
● Allow at-least-once production in regular clusters
● Some of the improvements are also useful for the non-at-least-once use cases

The benchmark
● Simulate the production traffic pattern (message size / # of partitions / QPS of each
topic, sample data etc.)
● Each broker leads ~ 900 topic-partitions
● ~ 100 K messages per broker (as leader)
● ~ 65 MB/s per broker (as leader)
● Snappy compression
● Each topic is consumed by 4 consumers

Testing cluster
11 servers with following configuration
CPU 2.40GHz 20 cores 40 hyperthreads
Memory 256GB DDR4 ECC in 4 channels
OS Linux Debian Jessie Kernel 4.4.78
JDK Oracle 1.8.0_65
RAID card 1GB write-back cache
HDD 22 disks in RAID-0, 40TB space in total
Kafka 10.0.2
Scala 2.10

The benchmark result before the optimizations
P99 latency 60 ms ~ 80 ms

How does at-least-once work in Kafka?
● Producers and consumers only talk to the leader broker
● Follower brokers keep fetching from the leader
Producer Consumer
Follower Broker
Leader Broker
Follower Broker
2. Fetch / 3. Response
2. Fetch / 3. Response
1. Produce /
4. Response

Improvement #1: Increase consumer fetch size
● Most consumers are now using the default fetch.min.bytes setting, which is 1
byte
● In our production environment, each topic is typically consumed by 4
consumers:
○ The actual user of the topic
○ Mirror Maker x 2 (for failover / backup)
○ Auditing
● Setting fetch.min.bytes to 64KB reduces the number of consumer fetch
requests from ~ 30K to ~5K
● The P99 latency is reduced by about 10ms, and P50 latency is reduced by
about 5ms

Improvement #2: Garbage collection
● Young gen GC happens about 1 once per second; each time, it stops-the-world
for 20ms to 30ms
● Larger young gen size helps reduce the GC overhead
● But, very large heap size (especially large old gen) has some negative impact
● When heap size < 32GB, Java uses 4-byte pointers instead of 8-byte pointers

Improvement #2: Garbage collection
From
-Xms36G -Xmx36G -XX:+UseG1GC -XX:MaxGCPauseMillis=20 -XX:NewRatio=2
To
-Xms22G -Xmx22G -XX:+UseG1GC -XX:NewSize=16G -XX:MaxNewSize=16G
-XX:InitiatingHeapOccupancyPercent=3
-XX:G1MixedGCCountTarget=1 -XX:G1HeapWastePercent=1
● P99 latency reduced by ~ 10ms, P50 latency reduced by ~ 0.5 to 1 ms

Improvement #3: Fix toString
● In Kafka broker code, some toString methods are declared like this:
case class FetchMetadata(...) {
override def toString =
"[minBytes: " + fetchMinBytes + ", " +
"onlyLeader:" + fetchOnlyLeader + ", "
"onlyCommitted: " + fetchOnlyCommitted + ", "
"partitionStatus: " + fetchPartitionStatus + "]"
}
● This String is only used in trace log
● This fix reduces P99 latency by a few ms
case class FetchMetadata(...) {
override def toString() = {
"[minBytes: " + fetchMinBytes + ", " +
"onlyLeader:" + fetchOnlyLeader + ", "
"onlyCommitted: " + fetchOnlyCommitted + ", "
"partitionStatus: " + fetchPartitionStatus + "]"
}
}

Improvement #4: Follower fetch protocol
Fetch Response (Version: 3) => {
throttle_time_ms => INT32
responses => [
topic => STRING
partition_responses => [
partition_header => {
partition => INT32
error_code => INT16
high_watermark => INT64
}
record_set => {RECORDS}
] * (# of partitions)
] * (# of topics)
}
Fetch Request (Version: 3) => {
replica_id => INT32
max_wait_time => INT32
min_bytes => INT32
max_bytes => INT32
topics => [
topic => STRING
partitions => [
partition => INT32
fetch_offset => INT64
max_bytes => INT32
] * (#of partitions)
] * (# of topics)
}

● Low producer latency = high follower fetch frequency
● In each fetch request and response, all topic-partitions are repeated, even
when there is no data
● In the benchmark, each broker is following 1800 topic-partitions

● The QPS of the production topic-partitions at Uber:
QPS range # of topic-partitions percentage
<0.01 18277 67.21
0.1 - 1 1607 5.91
1 - 10 1989 7.31
10 - 100 2230 8.20
100 - 1000 2565 9.43
> 1000 524 1.93

● Skipping empty topic-partitions in the follower fetch response
● Reduced P99 latency from ~35ms to ~30ms
● Reduced P50 latency from 8ms to 7ms

Improvement #5: Speed up (de)serialization
● Kafka generates ~2.6KB temporary objects for each topic-partition in each
fetch cycle, while the fetch protocol only needs about 100 bytes
● bytebuffer converted into a object tree, and then converted to a HashMap
● int, long are translated into Integer, Long

Improvement #5: Speed up (de)serialization(cont.)
● The solution:
○ Avoid generating the intermediate data structure
○ Use primitive types instead of boxed types
○ Cache topic names and topic-partition objects
● P99 latency is reduced by a few ms
● P50 latency is slightly improved

Improvement #6: Fewer threads
● All the network threads and IO threads in Kafka broker share one
ArrayBlockingQueue:
● The current production setting:
○ 32 network threads
○ 32 io threads

Improvement #6: Fewer threads
● Reducing the number of threads from 32 / 32 to 16x16
● P50 latency is reduced by about 1 ms
● Further improvement may require a different threading model or a more
efficient concurrent queue (e.g. disruptor? )

Improvement #7: Lock contention in purgatory
● Purgatory is the data structure for delayed operations (e.g. produce / fetch)
● Purgatory is protected with a read-write lock shared by all the io threads
● Each second, there are tens of thousands of operations added to and removed
from Purgatory
● The solution: sharding purgatory into 256 partitions
● P50 latency is reduced by another 1 ms

Improvement #8: I/O optimizations
● The P99 latency jumps a lot after several minutes
● This happens after the memory is used up by Linux page cache
● Disk usage also increases

● Most Kafka I/O operations are handled in memory by Linux page cache
● File system metadata changes are handled by the non-volatile memory on
RAID card
● Normally, Kafka threads should not be blocked by HDD I/O
● However, there are some exceptions ...

● There is an optional flush
● Write operations may have to load file system metadata (e.g. inode) and the last
page of the file from disk (I’m not an Linux expert, so not 100% sure about this)
● Rolling a segment file needs dentry and inode
● The solution:
○ Turn off flush
○ Periodically touch the last 2 pages of each index and segment file

● Kafka index lookup is needed in both consumer fetch and follower fetch. Normally, the followers and
consumers only look up the recent offsets. Binary search is not cache-friendly ...

● More cache-friendly binary search: look up only in the warm area, if possible

The overall performance improvement
● Benchmark
○ P99 latency: ~70 ms ⇨ less than 20 ms
○ P50 latency: ~15 ms ⇨ 3-4 ms
● Production at-least-once cluster
○ P99 latency: ~10 ms ⇨ 1 ms
○ P50 latency: ~5 ms ⇨ ~0.5 ms
● Production log cluster (acks=0)
○ CPU usage: ~50% ⇨ ~16%

Ideas for further improvements
● Refer topics with an integer ID instead of string
● Register followers / consumers on leader, instead of repeating all the topic
partitions in each fetch request
● The ArrayBlockingQueue between network threads and io threads is still a
bottleneck. We should consider either a different threading model or a more
efficient alternative, like disruptor
● Save data in distributed file system instead of local disk

Proprietary and confidential © 2017 Uber Technologies, Inc. All rights reserved. No part of this
document may be reproduced or utilized in any form or by any means, electronic or mechanical,
including photocopying, recording, or by any information storage or retrieval systems, without
permission in writing from Uber. This document is intended only for the use of the individual or entity
to whom it is addressed and contains information that is privileged, confidential or otherwise exempt
from disclosure under applicable law. All recipients of this document are notified that the information
contained herein includes proprietary and confidential information of Uber, and recipient may not
make use of, disseminate, or in any way disclose this document or any of the enclosed information
to any person other than employees of addressee to the extent necessary for consultations with
authorized personnel of Uber.

Improving Kafka at-least-once performance at Uber

More Related Content

What's hot (20)

Similar to Improving Kafka at-least-once performance at Uber (20)

Recently uploaded (20)

Improving Kafka at-least-once performance at Uber