Flink通过滚动窗口达到滑动窗口目的 节省内存和CPU资源(背压)

Flink在实时处理滑动窗口数据时, 由于窗口时间长, 滑动较为频繁, 导致算子计算压力过大, 下游算子计算速度抵不上上游数据产生速度, 会出现背压现象.

需求: 统计6小时用户设备共同用户数, 每10min统计一次

公共类

@Data
@AllArgsConstructor
// flatMap转换对象
private static class UserDevice {
    private final String userId;
    private final String deviceId;
}

@Data
// 用户设备统计结果
// 第一个map存放用户最新设备, 直接put覆盖, 取最新设备
// 第二个map存放设备对应用户, 因为要去重, 所以使用set存放
private static class UserDeviceSummary {
    private final Map<String, String> userDevices = new HashMap<>(60000); // (uid, did)
    private final Map<String, Set<String>> deviceUsers = new HashMap<>(60000); // (did, Set<uid>)
}

原算子 滑动窗口

dataStreamSource
    .flatMap((FlatMapFunction<JSONArray, UserDevice>) (array, collector) -> {
        try {
            array.forEach(e -> {
                JSONObject one = (JSONObject) e;
                // 只处理opay_show事件  app_name in ('opay', '1')
                if (one.containsKey("uid") && one.containsKey("did")) {
                    collector.collect(new UserDevice(one.getString("uid"), one.getString("did")));
                }
            });
        } catch (Exception ignored) {
        }
    }).returns(TypeInformation.of(new TypeHint<UserDevice>() {
    })).name("Stream flat map")
    .timeWindowAll(Time.hours(6), Time.minutes(10)) // 滑动窗口
    .allowedLateness(Time.minutes(1))
    .process(new ProcessAllWindowFunction<UserDevice, UserDeviceSummary, TimeWindow>() {
        @Override
        public void process(ProcessAllWindowFunction<UserDevice, UserDeviceSummary, TimeWindow>.Context context, Iterable<UserDevice> elements, Collector<UserDeviceSummary> out) throws Exception {
            UserDeviceSummary uds = new UserDeviceSummary();
            for (UserDevice ud : elements) {
                try {
                    // 不用线程安全集合, 提升效率 由于并行度为1, 应该不会有并发
                    uds.getUserDevices().put(ud.getUserId(), ud.getDeviceId());
                    if (!uds.getDeviceUsers().containsKey(ud.getDeviceId())) {
                        uds.getDeviceUsers().put(ud.getDeviceId(), new HashSet<>());
                    }
                    uds.getDeviceUsers().get(ud.getDeviceId()).add(ud.getUserId());
                } catch (Exception ignore) {
                }
            }
            out.collect(uds);
        }
    }).name("Process to Map")
    .process(new ProcessFunction<UserDeviceSummary, Map<String, Integer>>() {
        @Override
        public void processElement(UserDeviceSummary uds, ProcessFunction<UserDeviceSummary, Map<String, Integer>>.Context ctx, Collector<Map<String, Integer>> out) throws Exception {
            Map<String, Integer> result = new HashMap<>();
            for (String uid : uds.getUserDevices().keySet()) {
                try {
                    int count = uds.getDeviceUsers().get(uds.getUserDevices().get(uid)).size();
                    result.put(uid, count);
                } catch (Exception e) {
                    System.out.println("Process for sink error: " + e.getMessage());
                }
            }
            out.collect(result);
            // 清空数据 协助gc
            uds.getUserDevices().clear();
            uds.getDeviceUsers().clear();
            result.clear();
        }
    }).name("User device calc").print();

开始运行正常, 随着时间的推移, 数据堆积越来越大, 滑动过程中, 最大会有6h / 10min = 36次并行计算, cpu压力比较大, 并行度只能为1

优化

使用滚动窗口替换滑动窗口, 既节省了内存, 也减少了cpu计算. 每10min滚动一次, 外部使用queue存储, 最大保存36个元素

private static final int SUMMARY_LIST_CAPACITY = 36;
// merge list中36个元素 生成一个新的元素, 输出到下游
private static UserDeviceSummary merge(List<UserDeviceSummary> list) {
    UserDeviceSummary result = list.get(0);
    // 此处最好应该添加summary时间, 避免长时间没数据流入导致数据错误
    int length = Math.min(list.size(), SUMMARY_LIST_CAPACITY);
    System.out.println("Merge tumbling summary: " + length);
    for (int i = 1; i < length; i++) {
        UserDeviceSummary current = list.get(i);
        result.getUserDevices().putAll(current.getUserDevices());
        current.getDeviceUsers().forEach((key, value) -> result.getDeviceUsers().merge(key, value, (s1, s2) -> {
            s1.addAll(s2);
            return s1;
        }));
    }
    return result;
}

List<UserDeviceSummary> list = new LinkedList<>();

dataStreamSource
    .flatMap((FlatMapFunction<JSONArray, UserDevice>) (array, collector) -> {
        try {
            array.forEach(e -> {
                JSONObject one = (JSONObject) e;
                // 只处理opay_show事件  app_name in ('opay', '1')
                if (one.containsKey("uid") && one.containsKey("did")) {
                    collector.collect(new UserDevice(one.getString("uid"), one.getString("did")));
                }
            });
        } catch (Exception ignored) {
        }
    }).returns(TypeInformation.of(new TypeHint<UserDevice>() {
    })).name("Stream flat map")
    .timeWindowAll(Time.minutes(10)) // 使用滚动窗口代替滑动窗口, 节省资源
    .process(new ProcessAllWindowFunction<UserDevice, UserDeviceSummary, TimeWindow>() {
        @Override
        public void process(ProcessAllWindowFunction<UserDevice, UserDeviceSummary, TimeWindow>.Context context, Iterable<UserDevice> elements, Collector<UserDeviceSummary> out) throws Exception {
            UserDeviceSummary uds = new UserDeviceSummary();
            for (UserDevice ud : elements) {
                try {
                    // 不用线程安全集合, 提升效率
                    uds.getUserDevices().put(ud.getUserId(), ud.getDeviceId());
                    if (!uds.getDeviceUsers().containsKey(ud.getDeviceId())) {
                        uds.getDeviceUsers().put(ud.getDeviceId(), new HashSet<>());
                    }
                    uds.getDeviceUsers().get(ud.getDeviceId()).add(ud.getUserId());
                } catch (Exception ignore) {
                }
            }
            list.add(uds);
            if (list.size() > SUMMARY_LIST_CAPACITY) {
                list.remove(0);
            }
            out.collect(merge(list));
        }
    }).name("Process to Map")
    .process(new ProcessFunction<UserDeviceSummary, Map<String, Integer>>() {
        @Override
        public void processElement(UserDeviceSummary uds, ProcessFunction<UserDeviceSummary, Map<String, Integer>>.Context ctx, Collector<Map<String, Integer>> out) throws Exception {
            Map<String, Integer> result = new HashMap<>();
            for (String uid : uds.getUserDevices().keySet()) {
                try {
                    int count = uds.getDeviceUsers().get(uds.getUserDevices().get(uid)).size();
                    result.put(uid, count);
                } catch (Exception e) {
                    System.out.println("Process for sink error: " + e.getMessage());
                }
            }
            out.collect(result);
            uds.getUserDevices().clear();
            uds.getDeviceUsers().clear();
            result.clear();
        }
    }).name("User device calc").print();

再次部署, 服务运行正常!