每隔一段时间,删除N天前的数据,索引只保留最近几天的数据(索引不是按照日期生成的,不能直接删除整个索引)。【elasticsearch-version-5.x】
使用接口_delete_by_query,定期向集群提交批量删除任务,http请求不用等待删除任务完成才返回,而是在提交任务之后即时返回任务ID。使用_tasks接口定期检查删除任务的运行状态。这种方式解决了在删除大批量数据的时候Read timed out问题(_delete_by_query接口设置批量提交对于这个问题无解)。
在实际工程使用中,我们需要把elasticsearch的http接口全部封装为JavaWeb工程开发者易于使用和理解的依赖工程的形式。因此在下面的实现中保留此种方式,没有完全按照脚本的形式实现,而是通过jar+shell的形式实现这个功能,并且在封装的es接口包里面保留了这个删除接口。
# _delete_by_query接口
http://localhost:9210/indexName/indexType/_delete_by_query?refresh=true&scroll_size=1000&conflicts=proceed&wait_for_completion=false
# _tasks接口
http://localhost:9210/_tasks/EXlbuEGgRZK-IYKoOHmqWQ:990296121
#!/usr/bin/env bash
myJarPath=./lib/xxx.jar
# ---------------------------启动索引数据删除进程---------------------------
# 索引类型
indexType="indexType"
# 索引名称-多个索引名称使用逗号分隔
indexName="indexName"
# IP和端口-使用冒号分隔
ipPort="localhost:9200"
# 索引mapping中的时间字段
timeField="pubtime"
# 每隔delayTime执行一次删除数据操作 - 延时执行-支持按天/小时/分钟(格式数字加d/h/m:1d/24h/60m/60s)
delayTime="2s"
# 删除beforeDataTime以前的数据 - 行一次时删除多久以前的数据-支持按天/小时/分钟(格式数字加d/h/m:1d/24h/60m/60s)
beforeDataTime="2d"
# 是否启动DEBUG模式
debug="true"
#*****************************************************************
# 是否启用force merge(释放磁盘空间 - cpu/io消耗增加,缓存失效)
# 1、对于不再生成新分段的索引,建议打开此配置;2、如果索引在不断的产生新分段建议关闭此配置-通过修改集群段合并策略优化
#*****************************************************************
isForceMerge="false"
nohup java -Xmx512m -cp ${myJarPath} casia.isi.delete.DeleteIndexData ${indexType} ${indexName} ${ipPort} ${timeField} ${delayTime} ${beforeDataTime} ${debug} ${isForceMerge} >>logs/delete.DeleteIndexData.log 2>&1 &
package casia.isi.elasticsearch.operation.delete.shell;
/**
* ┏┓ ┏┓+ +
* ┏┛┻━━━━━━━┛┻┓ + +
* ┃ ┃
* ┃ ━ ┃ ++ + + +
* █████━█████ ┃+
* ┃ ┃ +
* ┃ ┻ ┃
* ┃ ┃ + +
* ┗━━┓ ┏━┛
* ┃ ┃
* ┃ ┃ + + + +
* ┃ ┃ Code is far away from bug with the animal protecting
* ┃ ┃ +
* ┃ ┃
* ┃ ┃ +
* ┃ ┗━━━┓ + +
* ┃ ┣┓
* ┃ ┏┛
* ┗┓┓┏━━━┳┓┏┛ + + + +
* ┃┫┫ ┃┫┫
* ┗┻┛ ┗┻┛+ + + +
*/
import casia.isi.elasticsearch.common.FieldOccurs;
import casia.isi.elasticsearch.common.RangeOccurs;
import casia.isi.elasticsearch.operation.delete.EsIndexDelete;
import casia.isi.elasticsearch.util.DateUtil;
import casia.isi.elasticsearch.util.StringUtil;
import com.alibaba.fastjson.JSONObject;
/**
* @Description: TODO(监控删除索引数据)
* @date 2019/5/30 15:27
*/
public final class DeleteDataByShell {
private static EsIndexDelete esIndexDataDelete;
private static String indexType;
private static String indexName;
private static String ipPort;
private static String timeField;
private static String delayTime;
private static String beforeDataTime;
private static boolean isForceMerge = false;
// DELETE WORK TASK ID
private static String lastTaskId;
public static boolean debug = false;
/**
* @param indexType:索引类型
* @param indexName:索引名称-多个索引名称使用逗号分隔
* @param ipPort:IP和端口-使用冒号分隔
* @param timeField:索引mapping中的时间字段
* @param delayTime:延时执行-支持按天/小时/分钟(格式数字加d/h/m:1d/24h/60m/60s)
* @param beforeDataTime:执行一次时删除多久以前的数据-支持按天/小时/分钟(格式数字加d/h/m:1d/24h/60m/60s)
* @param isForceMerge:true启用force-merge
* @return
* @Description: TODO(为监控程序创建一个索引数据删除对象)
*/
public DeleteDataByShell(String indexType, String indexName, String ipPort, String timeField,
String delayTime, String beforeDataTime, boolean isForceMerge) {
this.esIndexDataDelete = new EsIndexDelete(ipPort, indexName, indexType);
this.indexType = indexType;
this.indexName = indexName;
this.ipPort = ipPort;
this.timeField = timeField;
this.delayTime = delayTime;
this.beforeDataTime = beforeDataTime;
this.isForceMerge = isForceMerge;
}
/**
* @return
* @Description: TODO(启动监控删除)
*/
public void run() {
boolean isExcute = check();
while (isExcute) {
try {
// 执行删除
executeDelete();
// 延时执行
sleep();
} catch (Exception e) {
System.out.println("Delete data exception,please check your parameters!");
System.out.println("indexType:" + indexType);
System.out.println("indexName:" + indexName);
System.out.println("ipPort:" + ipPort);
System.out.println("timeField:" + timeField);
System.out.println("delayTime:" + delayTime);
System.out.println("beforeDataTime:" + beforeDataTime);
esIndexDataDelete.reset();
}
}
}
private boolean check() {
if (this.timeField != null && this.delayTime != null && this.beforeDataTime != null) {
return true;
}
return false;
}
private void sleep() throws InterruptedException {
Thread.sleep(dhmToMill(delayTime));
}
private void outputResult() {
System.out.println("Delay time:" + delayTime);
System.out.println("Delete data from " + beforeDataTime + " ago.Current system time:" + DateUtil.millToTimeStr(System.currentTimeMillis()));
if (debug) {
System.out.println("Query url:" + esIndexDataDelete.getQueryUrl());
System.out.println("Query json:" + esIndexDataDelete.getQueryString());
System.out.println("Query result json:" + esIndexDataDelete.getQueryReslut());
}
lastTaskId = setTaskId(esIndexDataDelete.getQueryReslut());
}
/**
* @param { "task": "EXlbuEGgRZK-IYKoOHmqWQ:xxxxxxx"
* }
* @return
* @Description: TODO(设置taskID)
*/
private String setTaskId(String queryReslut) {
JSONObject object = JSONObject.parseObject(queryReslut);
return object.getString("task");
}
private void executeDelete() {
// 输出上一个task的信息
System.out.println("===========================================EXECUTE DELETE TASK===========================================");
if (lastTaskId != null && !"".equals(lastTaskId)) {
System.out.println(esIndexDataDelete.outputLastTaskInfo(lastTaskId));
}
String currentThreadTime = getCurrentThreadTime();
esIndexDataDelete.addRangeTerms(timeField, currentThreadTime, FieldOccurs.MUST, RangeOccurs.LTE);
esIndexDataDelete.setRefresh(true);
esIndexDataDelete.setScrollSize(1000);
esIndexDataDelete.conflictsProceed("proceed");
esIndexDataDelete.setWaitForCompletion(false);
esIndexDataDelete.execute();
// 输出删除统计结果
outputResult();
// 释放磁盘空间(执行段合并操作)- CPU/IO消耗增加,缓存失效
if (isForceMerge) {
System.out.println(esIndexDataDelete.forceMerge());
}
esIndexDataDelete.reset();
}
private String getCurrentThreadTime() {
long mill = System.currentTimeMillis() - dhmToMill(beforeDataTime);
return DateUtil.millToTimeStr(mill);
}
private long dhmToMill(String dhmStr) {
if (dhmStr != null && !"".equals(dhmStr)) {
int number = Integer.valueOf(StringUtil.cutNumber(dhmStr));
if (dhmStr.contains("d")) {
return number * 86400000;
} else if (dhmStr.contains("h")) {
return number * 3600000;
} else if (dhmStr.contains("m")) {
return number * 60000;
} else if (dhmStr.contains("s")) {
return number * 1000;
}
}
return 0;
}
/**
* @param
* @return
* @Description: TODO(Delete thread main entrance)
*/
public static void main(String[] args) {
String indexType = args[0];
String indexName = args[1];
String ipPort = args[2];
String timeField = args[3];
String delayTime = args[4];
String beforeDataTime = args[5];
DeleteDataByShell.debug = Boolean.valueOf(args[6]);
String isForceMerge = args[7];
new DeleteDataByShell(indexType, indexName, ipPort, timeField, delayTime, beforeDataTime, Boolean.valueOf(isForceMerge)).run();
}
}
经过以上操作索引中的数据可以被正确的标记为删除,并且及时刷新查询显示。但是标记刷新之后,索引分段数据并没有将磁盘空间及时释放,还依赖于lucene分段合并的处理。
使用forcemerge可以及时释放磁盘空间,但是会带来cpu/io消耗增加,缓存失效等问题。这种问题对查询性能带来影响。但是可以按照具体的使用场景来采取措施:1、对于不再生成新分段的索引(不再有数据被索引和更新),可以考虑人工启动分段merge操作;2、如果索引在不断的产生新分段(数据被索引),通过修改集群段合并策略优化。在我们的需求中则必须采用第二种方式,线上系统人工_forcemerge带来的性能问题是不可接受的。
es默认每秒进行自动刷新,这带来的好处是新索引的数据可以及时对搜索可见。随之带来的问题是影响性能:某些缓存将会失效,拖慢搜索请求,而且重新打开索引的过程本身也需要一些处理能力,拖慢了索引的建立。
// 索引级setting
"index.refresh_interval": "5s",
flush操作是将内存数据冲刷到磁盘。内存缓冲区已满、事务日志已满、时间间隔已到,都会触发flush操作。具体策略请查阅相关文档。
// 集群配置elasticsearch.yml-内存缓冲区大小在elasticsearch.yml配置文件定义-可设置为JVM堆内存的百分比10%
"indices.memory.index_buffer_size":"3gb"
// 索引级setting-触动冲刷得规模-可设置为JVM堆内存得百分比10%(默认512mb)
"index.translog.flush_threshold_size": "3gb"
// 索引级setting-冲刷之间的时间间隔(默认是30m)
"index.translog.flush_threshold_period": "30m"
使用lucene默认的分层合并策略。关于分层合并策略的介绍请移步es官网。
// 索引级setting-每层分段数(segments_per_tier设为与max_merge_at_once相等可减少合并次数)
"index.merge.policy.segments_per_tier":5
// 索引级setting-每层合并的最大分段数(默认是10)
"index.merge.policy.max_merge_at_once": 5
// 索引级setting-最大分段规模(默认是5g)
"index.merge.policy.max_merged_segment": "1gb"
// 索引级setting-用于合并的最大线程数(设置为1可以让磁盘更好的运转)
// 要注意的是如果你是用HDD而非SSD的磁盘的话,最好是用单线程为妙。
"index.merge.scheduler.max_thread_count": 1
存储限流和存储的优化可以有效提升I/O的吞吐量。
存储限流的原因:过度的合并会拖慢集群。由于I/O的等待,会导致CPU负载也会很高。
// 集群配置elasticsearch.yml存储限流设置默认20mb(SSD-增加到100~200MB)
"indices.store.throttle.max_bytes_per_sec":"20mb"
// 集群配置elasticsearch.yml使存储限流的设置应用到所有的es操作
"indices.store.throttle.type":"all"
存储使用默认存储,主要考虑调整存储限流的设置。
存储类型:1、mmapfs-通常用于大型文件。eg.词条字典;2、niofs-其它类型文件。eg.存储字段。详细优化手段请移步es官方参考文档。
// URL
PUT http://localhost:9210/indexName/_settings
// PARAMETERS
{
"index.refresh_interval": "5s",
"index.translog": {
"flush_threshold_size": "3gb",
},
"index.merge": {
"policy": {
"segments_per_tier": 5,
"max_merge_at_once": 5,
"max_merged_segment": "1gb"
},
"scheduler.max_thread_count": 1
}
}
// RESPONSBODY
{
"acknowledged": true
}
// 使用GET接口查看setting
GET http://localhost:9210/indexName/_settings
{
"indexName": {
"settings": {
"index": {
"refresh_interval": "5s",
"number_of_shards": "5",
"translog": {
"flush_threshold_size": "3gb"
},
"provided_name": "indexName",
"merge": {
"scheduler": {
"max_thread_count": "1"
},
"policy": {
"segments_per_tier": "5",
"max_merge_at_once": "5",
"max_merged_segment": "1gb"
}
},
"creation_date": "1559195227068",
"number_of_replicas": "0",
"uuid": "aDekoukTQL2HeB_aQy_HFA",
"version": {
"created": "5060399"
}
}
}
}
}
postman设置index的setting:
Lucene分段处理优化之后,很明显可以看到Heap Memory消耗下降了将近一般左右(之前的图有一个驼峰式的下降效果忘记截图了:)gg):
使用_tasks接口,计算平均处理速率。
http://localhost:9210/_tasks/EXlbuEGgRZK-IYKoOHmqWQ:98453352X
{
"completed": true,
"task": {
"node": "EXlbuEGgRZK-IYKoOHmqWQ",
"id": 984533525,
"type": "transport",
"action": "indices:data/write/delete/byquery",
"status": {
"total": 10399385,
"updated": 0,
"created": 0,
"deleted": 4784168,
"batches": 10400,
"version_conflicts": 5615217,
"noops": 0,
"retries": {
"bulk": 0,
"search": 0
},
"throttled_millis": 0,
"requests_per_second": -1,
"throttled_until_millis": 0
},
"description": "delete-by-query [indexName]",
"start_time_in_millis": 1559727929590,
"running_time_in_nanos": 3237112234217,
"cancellable": true
},
"response": {
"took": 3237112,
"timed_out": false,
"total": 10399385,
"updated": 0,
"created": 0,
"deleted": 4784168,
"batches": 10400,
"version_conflicts": 5615217,
"noops": 0,
"retries": {
"bulk": 0,
"search": 0
},
"throttled_millis": 0,
"requests_per_second": -1,
"throttled_until_millis": 0,
"failures": []
}
}
类似上述结果,可以根据task的运行情况计算处理效率。使用running_time_in_nanos和deleted字段的数据计算平均处理速率。服务器配置:1、Intel® Xeon® CPU E5-2620 v4 @ 2.10GHz-32核,2、磁盘-HDD1.6T,3、内存-128G。
数据量/总耗时 | 速率 |
---|---|
100万/792s/13分钟 | 1262t/s |
219万/1768s/29分钟 | 1238t/s |
480万/3237s/53分钟 | 1482t/s |
在如上的task统计结果中,可以看到有很多数据是标记为version_conflicts。在轮询的删除过程中需要被删除的数据最终都会被删除(每30分钟运行一次删除进程)。如果对于数据删除时效性要求比较高的话,需要解决这个问题。并且继续优化删除策略。
// 没有数据版本冲突的删除任务,返回的信息是这样的(version_conflicts=0)
{
"completed": true,
"task": {
"node": "EXlbuEGgRZK-IYKoOHmqWQ",
"id": 990296121,
"type": "transport",
"action": "indices:data/write/delete/byquery",
"status": {
"total": 170733,
"updated": 0,
"created": 0,
"deleted": 170733,
"batches": 171,
"version_conflicts": 0,
"noops": 0,
"retries": {
"bulk": 0,
"search": 0
},
"throttled_millis": 0,
"requests_per_second": -1,
"throttled_until_millis": 0
},
"description": "delete-by-query [news_small, blog_small, forum_threads_small, mblog_info_small, video_brief_small, wechat_message_xigua_small, appdata_small, newspaper_info_small][monitor_caiji_small]",
"start_time_in_millis": 1559731529771,
"running_time_in_nanos": 71981947551,
"cancellable": true
},
"response": {
"took": 71981,
"timed_out": false,
"total": 170733,
"updated": 0,
"created": 0,
"deleted": 170733,
"batches": 171,
"version_conflicts": 0,
"noops": 0,
"retries": {
"bulk": 0,
"search": 0
},
"throttled_millis": 0,
"requests_per_second": -1,
"throttled_until_millis": 0,
"failures": []
}
}
在调用_delete_by_query接口时,设置参数refresh=wait_for。
refresh参数-true表示:立即刷新主分片和副分片;false:表示不刷新,不设置此条件默认不刷新;wait_for:使用集群自动刷新机制(默认1s,在索引级自定义5s或者其它值,根据业务决定。本次测试使用的5s)。
经过_tasks接口统计,发现优化这个参数之后,每秒的处理能力提升了3~4倍,1262t/s->4115t/s。
数据量/总耗时 | 速率 |
---|---|
100万/243s/4分钟 | 4115t/s |
122万/297s/5分钟 | 4107t/s |
{
"completed": true,
"task": {
"node": "EXlbuEGgRZK-IYKoOHmqWQ",
"id": 1111458358,
"type": "transport",
"action": "indices:data/write/delete/byquery",
"status": {
"total": 1215333,
"updated": 0,
"created": 0,
"deleted": 1215333,
"batches": 1216,
"version_conflicts": 0,
"noops": 0,
"retries": {
"bulk": 0,
"search": 0
},
"throttled_millis": 0,
"requests_per_second": -1,
"throttled_until_millis": 0
},
"description": "delete-by-query [indexName]",
"start_time_in_millis": 1559802968421,
"running_time_in_nanos": 297299330904,
"cancellable": true
},
"response": {
"took": 297299,
"timed_out": false,
"total": 1215333,
"updated": 0,
"created": 0,
"deleted": 1215333,
"batches": 1216,
"version_conflicts": 0,
"noops": 0,
"retries": {
"bulk": 0,
"search": 0
},
"throttled_millis": 0,
"requests_per_second": -1,
"throttled_until_millis": 0,
"failures": []
}
}