工作纪实51-手撸AB实验分流策略

前几天写了一篇关于哈希算法的文章，起源就是在构思AB实验平台的时候，用到了哈希，所以对其做了深入的了解
AB实验平台是一般互联网做策略、样式实验会用到的一个系统，一般开启某个实验之后，需要对线上流量进行分流：客户端->实验平台->策略平台->应用服务，大概是这个链路

场景

线上流量策略分流，需要对不同人群做策略实验，最终效果好的一组需要推全到100%，在未推全之前，100%的流量会被分到原有流量+几个实验组；
比如我有一个实验：样式迭代，产品希望60%的流量走线上逻辑，剩下的40%，均分到4个实验组，这样可以细致的比对样式的效果，当然，如果一个用户命中了实验组A，那么他后续的行为，在实验组不调整的情况下，需要一直保持（有的人说希望调整了也保持，但是真的了解AB分流逻辑之后，真挺难的）；
其实我们有自己的AB平台，但是我在想如果让我去实现，我会怎么做，便有了这一篇文章；

考虑的问题

1.如何保证一个用户对实验命中唯一性

在实验组不调整的情况下，PM希望用户分流的结果始终保持一致；
【策略】：hash算法，通过对imei进行哈希运算得到int的哈希值即可，可以加一个时间戳（实验修改时间），来控制用户的imei命中变化

2.如何按比例分流

【策略】取模运算，构建一个长度为100的bucket，然后通过对imei的哈希值运算结果对100取模，即可得到0-100的数，从而去命中数据区间，挑选实验组即可；由于哈希算法的特性，只要imei不变，时间戳(因子)不变，计算得到的哈希值是永远不会变的，所以取模结果，命中实验的结果也是不变的；

比如我的case是线上流量60%，其他实验组10%,10%,10%,10%，则可以构建一个数据区间：【0,60,70,80,90,100】

3.接口效率得有保证

• 不要对分流的结果用户存redis，开100个实验机器缓存能被打爆
• 分流计算要高效，极致的高效，要多留时间给业务侧

实战

代码

@Data
public class TrafficDistributor {
    private String id;// 实验id
    private String name;
    private double isolation;// 线上流量
    private List<Experiment> experiments;// 实验组
    private Table<String, String, Set<String>> whiteTable = HashBasedTable.create();
    private static Date updateTime = new Date(1719844946L);//2024/7/1 22:42:26

    public TrafficDistributor(String id, String name, double isolation, List<Experiment> experiments, Table<String, String, Set<String>> whiteTable) {
        this.id = id;
        this.name = name;
        this.isolation = isolation;
        this.experiments = experiments;
        this.whiteTable = whiteTable;
    }

    public static void main(String[] args) {
        // 实验id
        String id = "new_style";
        // 1.初始化白名单==》正式环境从db中拿出来初始化
        Table<String, String, Set<String>> white = HashBasedTable.create();
        white.put(id, "00", Sets.newHashSetWithExpectedSize(0));
        white.put(id, "01", Set.of("whitelist_imei_1", "whitelist_imei_2"));
        white.put(id, "02", Set.of("whitelist_imei_3", "whitelist_imei_4"));
        white.put(id, "11", Set.of("whitelist_imei_5", "whitelist_imei_6"));
        white.put(id, "12", Set.of("whitelist_imei_7", "whitelist_imei_8"));
        // 2.初始化实验组
        Experiment experiment1 = new Experiment("01", "对照组01", 10);
        Experiment experiment2 = new Experiment("02", "对照组02", 10);
        Experiment experiment3 = new Experiment("11", "实验组01", 10);
        Experiment experiment4 = new Experiment("12", "实验组02", 10);
        List<Experiment> experiments = List.of(experiment1, experiment2, experiment3, experiment4);
        // 3.初始化分流器
        TrafficDistributor distributor = new TrafficDistributor(id, "新样式实验", 60, experiments, white);
        // 4.模拟分流
        for (int times = 0; times < 100; times++) {
            int index = 1, size = 100;
            Map<String, Integer> result = Maps.newHashMap();
            List<Long> cost = Lists.newArrayListWithCapacity(size);
            while (index <= size) {
                String imei = RandomStringUtils.randomAlphanumeric(64);
                long start = System.currentTimeMillis();
                // 核心分流
                Experiment experiment = distributor.allocate(id, imei);
                if (experiment != null) {
                    // 实验组
                    result.compute(experiment.getId(), (eid, count) -> (count == null) ? 1 : count + 1);
                } else {
                    // 线上
                    int value = result.getOrDefault("00", 0);
                    result.put("00", value + 1);
                }
                cost.add(System.currentTimeMillis() - start);
                index++;
            }
            System.out.println("耗时:" + cost.stream().mapToLong(Long::longValue).sum() + "；结果:" + JSON.toJSONString(result));
        }
    }

    /**
     * 流量分配: 白名单 - 实验组 - 线上流量
     */
    public Experiment allocate(String id, String imei) {
        Experiment whiteListShot = whiteListShot(id, imei);
        if (whiteListShot != null) {
            return whiteListShot;
        }
        List<Double> weights = experiments.stream().map(Experiment::getWeight).collect(Collectors.toList());
        int groupId = distributeTraffic(weights, imei);
        if (groupId < 0) {
            // 线上流量
            return null;
        }
        // 实验组流量
        return experiments.get(groupId);
    }

    /**
     * 根据每个实验组的权重配比，判断最终流量应该分配到哪个实验组。
     *
     * @param weights 每个实验组的权重值数组。
     * @param imei    imei
     * @return 分配流量的实验组索引
     */
    public int distributeTraffic(List<Double> weights, String imei) {
        if (CollectionUtils.isEmpty(weights)) {
            return -1;
        }
        double totalWeight = 100;// 总权重100%
        double testWeight = weights.stream().mapToDouble(Double::doubleValue).sum();// 实验组总权重
        if (totalWeight != (isolation + testWeight)) {
            throw new IllegalArgumentException("实验组和对照组流量分配存在问题");
        }
        // imei+时间戳(因子)生成hash
        int hash = HashUtils.hashcode(imei + updateTime.getTime());
        return bucketShot(weights, hash % totalWeight);
    }

    /**
     * 哈希值进行取模定位后，命中的实验
     */
    public int bucketShot(List<Double> weights, double bucketIndex) {
        List<Double> range = Lists.newArrayListWithCapacity(weights.size() + 1);
        range.add(isolation);
        double pre = range.get(0);
        for (double weight : weights) {
            range.add(pre + weight);
            pre = pre + weight;
        }
        for (int i = 0; i < range.size(); i++) {
            if (bucketIndex <= range.get(i)) {
                return i - 1;
            }
        }
        throw new IllegalArgumentException("实验组和对照组流量分配存在问题");
    }

    /**
     * 白名单命中
     *
     * @param id   实验id
     * @param imei imei
     * @return 实验组
     */
    public Experiment whiteListShot(String id, String imei) {
        assert whiteTable != null && !whiteTable.isEmpty();
        if (!whiteTable.containsRow(id)) {
            throw new IllegalArgumentException("实验id=" + id + "不存在");
        }
        Map<String, Set<String>> experimentData = whiteTable.row(id);
        for (Map.Entry<String, Set<String>> entry : experimentData.entrySet()) {
            Set<String> values = entry.getValue();
            if (!CollectionUtils.isEmpty(values) && values.contains(imei)) {
                String key = entry.getKey();
                return experiments.stream().filter(test -> test.getId().equals(key)).findFirst().orElse(null);
            }
        }
        return null;
    }

    // 其他方法保持不变
    @Data
    @AllArgsConstructor
    static class Experiment {
        private String id;
        private String name;
        private double weight;
    }

public class HashUtils {
    /**
     * hashCode方法
     */
    public static int hashcode(Object obj) {
        final int p = 16777619;
        int hash = (int) 2166136261L;
        String str = obj.toString();
        for (int i = 0; i < str.length(); i++)
            hash = (hash ^ str.charAt(i)) * p;
        hash += hash << 13;
        hash ^= hash >> 7;
        hash += hash << 3;
        hash ^= hash >> 17;
        hash += hash << 5;

        if (hash < 0)
            hash = Math.abs(hash);
        return hash;
    }
}

分析

耗时:34；结果:{"11":11,"00":61,"01":7,"12":9,"02":12}
耗时:4；结果:{"11":8,"00":61,"12":4,"01":9,"02":18}
耗时:2；结果:{"11":9,"00":60,"01":10,"12":7,"02":14}
耗时:2；结果:{"11":11,"00":54,"01":10,"12":14,"02":11}
耗时:1；结果:{"11":9,"00":65,"12":12,"01":8,"02":6}
耗时:3；结果:{"11":9,"00":61,"01":6,"12":12,"02":12}
耗时:2；结果:{"11":9,"00":56,"12":8,"01":17,"02":10}
耗时:2；结果:{"11":7,"00":61,"12":7,"01":10,"02":15}
耗时:1；结果:{"11":8,"00":58,"01":13,"12":4,"02":17}
耗时:0；结果:{"11":5,"00":72,"01":9,"12":5,"02":9}
耗时:0；结果:{"11":2,"00":70,"12":14,"01":8,"02":6}
耗时:2；结果:{"11":7,"00":63,"12":10,"01":10,"02":10}
耗时:0；结果:{"11":8,"00":60,"12":11,"01":10,"02":11}
耗时:1；结果:{"11":9,"00":60,"12":11,"01":13,"02":7}
耗时:0；结果:{"11":12,"00":71,"12":5,"01":9,"02":3}
耗时:1；结果:{"11":12,"00":57,"01":11,"12":11,"02":9}
耗时:0；结果:{"11":8,"00":62,"01":14,"12":7,"02":9}
耗时:1；结果:{"11":10,"00":64,"12":6,"01":9,"02":11}
耗时:2；结果:{"11":5,"00":73,"12":7,"01":9,"02":6}
耗时:0；结果:{"11":9,"00":68,"01":6,"12":9,"02":8}
耗时:1；结果:{"11":12,"00":63,"01":10,"12":4,"02":11}
耗时:1；结果:{"11":15,"00":59,"01":8,"12":9,"02":9}
耗时:0；结果:{"11":10,"00":66,"01":8,"12":6,"02":10}
耗时:1；结果:{"11":8,"00":64,"01":10,"12":6,"02":12}
耗时:2；结果:{"11":11,"00":63,"01":8,"12":8,"02":10}
耗时:1；结果:{"11":5,"00":66,"12":9,"01":12,"02":8}
耗时:3；结果:{"11":8,"00":67,"12":10,"01":9,"02":6}
耗时:1；结果:{"11":9,"00":54,"12":16,"01":7,"02":14}
耗时:0；结果:{"11":11,"00":63,"12":5,"01":11,"02":10}
耗时:1；结果:{"11":10,"00":59,"01":12,"12":8,"02":11}
耗时:1；结果:{"11":12,"00":62,"12":11,"01":8,"02":7}
耗时:0；结果:{"11":8,"00":59,"12":8,"01":13,"02":12}
耗时:1；结果:{"11":12,"00":51,"12":11,"01":15,"02":11}
耗时:1；结果:{"11":1,"00":72,"01":10,"12":8,"02":9}
耗时:1；结果:{"11":8,"00":55,"01":18,"12":9,"02":10}
耗时:0；结果:{"11":6,"00":54,"01":22,"12":5,"02":13}
耗时:1；结果:{"11":9,"00":58,"12":11,"01":11,"02":11}
耗时:1；结果:{"11":15,"00":57,"12":12,"01":3,"02":13}
耗时:1；结果:{"11":6,"00":66,"12":10,"01":11,"02":7}
耗时:0；结果:{"11":13,"00":61,"12":12,"01":8,"02":6}
耗时:0；结果:{"11":8,"00":61,"12":11,"01":10,"02":10}
耗时:0；结果:{"11":10,"00":57,"01":10,"12":12,"02":11}
耗时:1；结果:{"11":6,"00":62,"12":12,"01":11,"02":9}
耗时:5；结果:{"11":9,"00":64,"12":8,"01":10,"02":9}
耗时:0；结果:{"11":15,"00":54,"12":8,"01":9,"02":14}
耗时:0；结果:{"11":12,"00":62,"01":8,"12":10,"02":8}
耗时:0；结果:{"11":9,"00":63,"12":6,"01":12,"02":10}
耗时:0；结果:{"11":9,"00":59,"01":10,"12":9,"02":13}
耗时:1；结果:{"11":10,"00":58,"01":7,"12":14,"02":11}
耗时:1；结果:{"11":9,"00":73,"01":8,"12":2,"02":8}
耗时:0；结果:{"11":14,"00":62,"12":7,"01":10,"02":7}
耗时:1；结果:{"11":12,"00":55,"12":10,"01":12,"02":11}
耗时:0；结果:{"11":5,"00":59,"01":7,"12":17,"02":12}
耗时:0；结果:{"11":10,"00":59,"12":7,"01":10,"02":14}
耗时:1；结果:{"11":11,"00":54,"01":17,"12":8,"02":10}
耗时:0；结果:{"11":8,"00":65,"12":8,"01":9,"02":10}
耗时:1；结果:{"11":13,"00":61,"12":8,"01":9,"02":9}
耗时:1；结果:{"11":6,"00":67,"01":10,"12":11,"02":6}
耗时:1；结果:{"11":7,"00":61,"12":8,"01":10,"02":14}
耗时:0；结果:{"11":6,"00":63,"12":8,"01":10,"02":13}
耗时:0；结果:{"11":9,"00":62,"12":9,"01":9,"02":11}
耗时:1；结果:{"11":5,"00":65,"01":8,"12":11,"02":11}
耗时:0；结果:{"11":11,"00":52,"12":9,"01":15,"02":13}
耗时:0；结果:{"11":14,"00":66,"01":4,"12":7,"02":9}
耗时:0；结果:{"11":12,"00":54,"12":8,"01":8,"02":18}
耗时:1；结果:{"11":9,"00":64,"12":8,"01":10,"02":9}
耗时:1；结果:{"11":9,"00":65,"01":3,"12":11,"02":12}
耗时:0；结果:{"11":5,"00":67,"01":7,"12":12,"02":9}
耗时:0；结果:{"11":13,"00":50,"01":12,"12":11,"02":14}
耗时:1；结果:{"11":18,"00":55,"12":7,"01":10,"02":10}
耗时:0；结果:{"11":5,"00":64,"01":12,"12":5,"02":14}
耗时:0；结果:{"11":10,"00":68,"12":6,"01":7,"02":9}
耗时:1；结果:{"11":9,"00":71,"12":4,"01":6,"02":10}
耗时:0；结果:{"11":8,"00":62,"12":9,"01":9,"02":12}
耗时:0；结果:{"11":10,"00":64,"12":8,"01":9,"02":9}
耗时:0；结果:{"11":9,"00":57,"12":10,"01":9,"02":15}
耗时:0；结果:{"11":10,"00":60,"12":13,"01":9,"02":8}
耗时:0；结果:{"11":12,"00":66,"01":5,"12":9,"02":8}
耗时:0；结果:{"11":6,"00":58,"01":11,"12":13,"02":12}
耗时:1；结果:{"11":10,"00":62,"01":12,"12":9,"02":7}
耗时:1；结果:{"11":7,"00":66,"12":11,"01":7,"02":9}
耗时:0；结果:{"11":10,"00":63,"12":9,"01":11,"02":7}
耗时:1；结果:{"11":8,"00":61,"12":10,"01":12,"02":9}
耗时:0；结果:{"11":8,"00":62,"12":6,"01":10,"02":14}
耗时:0；结果:{"11":7,"00":68,"12":8,"01":11,"02":6}
耗时:0；结果:{"11":11,"00":54,"01":11,"12":16,"02":8}
耗时:0；结果:{"11":7,"00":68,"01":10,"12":6,"02":9}
耗时:0；结果:{"11":7,"00":65,"12":7,"01":8,"02":13}
耗时:0；结果:{"11":8,"00":69,"01":8,"12":5,"02":10}
耗时:0；结果:{"11":15,"00":60,"01":6,"12":11,"02":8}
耗时:0；结果:{"11":9,"00":70,"01":6,"12":7,"02":8}
耗时:0；结果:{"11":14,"00":62,"12":7,"01":10,"02":7}
耗时:0；结果:{"11":11,"00":64,"12":7,"01":7,"02":11}
耗时:1；结果:{"11":6,"00":56,"12":14,"01":10,"02":14}
耗时:0；结果:{"11":7,"00":64,"12":8,"01":11,"02":10}
耗时:1；结果:{"11":14,"00":65,"12":4,"01":10,"02":7}
耗时:0；结果:{"11":13,"00":59,"12":7,"01":13,"02":8}
耗时:1；结果:{"11":5,"00":65,"12":8,"01":14,"02":8}
耗时:1；结果:{"11":12,"00":54,"01":11,"12":10,"02":13}
耗时:1；结果:{"11":10,"00":56,"12":11,"01":11,"02":12}
Process finished with exit code 0

• 我手写的hashCode方法，包括随机字符串生成的imei仍然存在实验组分类偏向的情况，中间尝试过使用一致性哈希解决偏移，但是又解决不了加权按比例分配的问题，不知道各位同仁有没有什么好的建议；我怀疑是随机字符串的偏向性问题，后面看看有没有办法解决【扩大了一下样本数1w+，发现差距小了很多】
• 取模算法是个宝藏，既可以精准定位，也可以利用随机性，来做区间命中
• updateTime它是个变化因子，如果实验有新增实验组或者流量调整，可以利用它来控制imei实验组变化

参考资料

• https://zhuanlan.zhihu.com/p/404232432
• https://blog.csdn.net/SmartCodeTech/article/details/113698568?spm=1001.2101.3001.6650.2&utm_medium=distribute.pc_relevant.none-task-blog-2%7Edefault%7EBlogCommendFromBaidu%7ECtr-2-113698568-blog-131205090.235%5Ev43%5Epc_blog_bottom_relevance_base8&depth_1-utm_source=distribute.pc_relevant.none-task-blog-2%7Edefault%7EBlogCommendFromBaidu%7ECtr-2-113698568-blog-131205090.235%5Ev43%5Epc_blog_bottom_relevance_base8&utm_relevant_index=5