深入了解如何基于Python读写Kafka

这篇文章主要介绍了深入了解如何基于Python读写Kafka,文中通过示例代码介绍的非常详细，对大家的学习或者工作具有一定的参考学习价值,需要的朋友可以参考下

本篇会给出如何使用python来读写kafka, 包含生产者和消费者.

以下使用kafka-python客户端

生产者

爬虫大多时候作为消息的发送端, 在消息发出去后最好能记录消息被发送到了哪个分区, offset是多少, 这些记录在很多情况下可以帮助快速定位问题, 所以需要在send方法后加入callback函数, 包括成功和失败的处理

# -*- coding: utf-8 -*-

'''
callback也是保证分区有序的, 比如2条消息, a先发送, b后发送, 对于同一个分区, 那么会先回调a的callback, 再回调b的callback
'''

import json
from kafka import KafkaProducer

topic = 'demo'

def on_send_success(record_metadata):
 print(record_metadata.topic)
 print(record_metadata.partition)
 print(record_metadata.offset)

def on_send_error(excp):
 print('I am an errback: {}'.format(excp))

def main():
 producer = KafkaProducer(
   bootstrap_servers='localhost:9092'
 )
 producer.send(topic, value=b'{"test_msg":"hello world"}').add_callback(on_send_success).add_callback(
   on_send_error)
 # close() 方法会阻塞等待之前所有的发送请求完成后再关闭 KafkaProducer
 producer.close()

def main2():
 '''
 发送json格式消息
 :return:
 '''
 producer = KafkaProducer(
   bootstrap_servers='localhost:9092',
   value_serializer=lambda m: json.dumps(m).encode('utf-8')
 )
 producer.send(topic, value={"test_msg": "hello world"}).add_callback(on_send_success).add_callback(
   on_send_error)
 # close() 方法会阻塞等待之前所有的发送请求完成后再关闭 KafkaProducer
 producer.close()
if __name__ == '__main__':
 # main()
 main2()

消费者

kafka的消费模型比较复杂, 我会分以下几种情况来进行说明

1.不使用消费组(group_id=None)

不使用消费组的情况下可以启动很多个消费者, 不再受限于分区数, 即使消费者数量 > 分区数, 每个消费者也都可以收到消息

# -*- coding: utf-8 -*-

'''
消费者: group_id=None
'''
from kafka import KafkaConsumer
topic = 'demo'
def main():
 consumer = KafkaConsumer(
   topic,
   bootstrap_servers='localhost:9092',
   auto_offset_reset='latest',
   # auto_offset_reset='earliest',
 )
 for msg in consumer:
   print(msg)
   print(msg.value)
 consumer.close()
if __name__ == '__main__':
 main()

2.指定消费组

以下使用pool方法来拉取消息

pool 每次拉取只能拉取一个分区的消息, 比如有2个分区1个consumer, 那么会拉取2次

pool 是如果有消息马上进行拉取, 如果timeout_ms内没有新消息则返回空dict, 所以可能出现某次拉取了1条消息, 某次拉取了max_records条

# -*- coding: utf-8 -*-

'''
消费者: 指定group_id
'''

from kafka import KafkaConsumer

topic = 'demo'
group_id = 'test_id'

def main():
 consumer = KafkaConsumer(
   topic,
   bootstrap_servers='localhost:9092',
   auto_offset_reset='latest',
   group_id=group_id,

)
 while True:
   try:
     # return a dict
     batch_msgs = consumer.poll(timeout_ms=1000, max_records=2)
     if not batch_msgs:
       continue
     '''
     {TopicPartition(topic='demo', partition=0): [ConsumerRecord(topic='demo', partition=0, offset=42, timestamp=1576425111411, timestamp_type=0, key=None, value=b'74', headers=[], checksum=None, serialized_key_size=-1, serialized_value_size=2, serialized_header_size=-1)]}
     '''
     for tp, msgs in batch_msgs.items():
       print('topic: {}, partition: {} receive length: '.format(tp.topic, tp.partition, len(msgs)))
       for msg in msgs:
         print(msg.value)
   except KeyboardInterrupt:
     break

consumer.close()

if __name__ == '__main__':
 main()

关于消费组

我们根据配置参数分为以下几种情况

• group_id=None

• auto_offset_reset='latest': 每次启动都会从最新出开始消费, 重启后会丢失重启过程中的数据

• auto_offset_reset='latest': 每次从最新的开始消费, 不会管哪些任务还没有消费

• 指定group_id

• auto_offset_reset='latest': 从上次提交offset的地方开始消费

• auto_offset_reset='earliest': 从上次提交offset的地方开始消费

• auto_offset_reset='latest': 只消费启动后的收到的数据, 重启后会从上次提交offset的地方开始消费

• auto_offset_reset='earliest': 从最开始消费全量数据

• 全新group_id

• 旧group_id(即kafka集群中还保留着该group_id的提交记录)

性能测试

以下是在本地进行的测试, 如果要在线上使用kakfa, 建议提前进行性能测试

producer

# -*- coding: utf-8 -*-

'''
producer performance

environment:
 mac
 python3.7
 broker 1
 partition 2
'''

import json
import time
from kafka import KafkaProducer

topic = 'demo'
nums = 1000000

def main():
 producer = KafkaProducer(
   bootstrap_servers='localhost:9092',
   value_serializer=lambda m: json.dumps(m).encode('utf-8')
 )
 st = time.time()
 cnt = 0
 for _ in range(nums):
   producer.send(topic, value=_)
   cnt += 1
   if cnt ％ 10000 == 0:
     print(cnt)

producer.flush()

et = time.time()
 cost_time = et - st
 print('send nums: {}, cost time: {}, rate: {}/s'.format(nums, cost_time, nums // cost_time))

if __name__ == '__main__':
 main()

'''
send nums: 1000000, cost time: 61.89236712455749, rate: 16157.0/s
send nums: 1000000, cost time: 61.29534196853638, rate: 16314.0/s
'''

consumer

# -*- coding: utf-8 -*-

'''
consumer performance
'''

import time
from kafka import KafkaConsumer

topic = 'demo'
group_id = 'test_id'

def main1():
 nums = 0
 st = time.time()

consumer = KafkaConsumer(
   topic,
   bootstrap_servers='localhost:9092',
   auto_offset_reset='latest',
   group_id=group_id
 )
 for msg in consumer:
   nums += 1
   if nums >= 500000:
     break
 consumer.close()

et = time.time()
 cost_time = et - st
 print('one_by_one: consume nums: {}, cost time: {}, rate: {}/s'.format(nums, cost_time, nums // cost_time))

def main2():
 nums = 0
 st = time.time()

consumer = KafkaConsumer(
   topic,
   bootstrap_servers='localhost:9092',
   auto_offset_reset='latest',
   group_id=group_id
 )
 running = True
 batch_pool_nums = 1
 while running:
   batch_msgs = consumer.poll(timeout_ms=1000, max_records=batch_pool_nums)
   if not batch_msgs:
     continue
   for tp, msgs in batch_msgs.items():
     nums += len(msgs)
     if nums >= 500000:
       running = False
       break

consumer.close()

et = time.time()
 cost_time = et - st
 print('batch_pool: max_records: {} consume nums: {}, cost time: {}, rate: {}/s'.format(batch_pool_nums, nums,
                                             cost_time,
                                             nums // cost_time))

if __name__ == '__main__':
 # main1()
 main2()

'''
one_by_one: consume nums: 500000, cost time: 8.018627166748047, rate: 62354.0/s
one_by_one: consume nums: 500000, cost time: 7.698841094970703, rate: 64944.0/s

batch_pool: max_records: 1 consume nums: 500000, cost time: 17.975456953048706, rate: 27815.0/s
batch_pool: max_records: 1 consume nums: 500000, cost time: 16.711708784103394, rate: 29919.0/s

batch_pool: max_records: 500 consume nums: 500369, cost time: 6.654940843582153, rate: 75187.0/s
batch_pool: max_records: 500 consume nums: 500183, cost time: 6.854053258895874, rate: 72976.0/s

batch_pool: max_records: 1000 consume nums: 500485, cost time: 6.504687070846558, rate: 76942.0/s
batch_pool: max_records: 1000 consume nums: 500775, cost time: 7.047331809997559, rate: 71058.0/s
'''

来源：https://www.cnblogs.com/zlone/p/12116817.html