让我们扩展 domains 以获得更好的覆盖范围：
domains = spark.createDataFrame([
"something.google.com", # OK
"something.google.com.somethingelse.ac.uk", # NOT OK
"something.good.com.cy", # OK
"something.good.com.cy.mal.org", # NOT OK
"something.bad.com.cy", # NOT OK
"omgalsogood.com.cy", # NOT OK
"good.com.cy", # OK
"sogood.example.com", # OK Match for shorter redundant, mismatch on longer
"notsoreal.googleecom" # NOT OK
], "string").toDF('domains')
good_domains = spark.createDataFrame([
"google.com", "good.com.cy", "alsogood.com.cy",
"good.example.com", "example.com" # Redundant case
], "string").toDF('gooddomains')
现在... 一个简单的解决方案，只使用 Spark SQL 原语，就是稍微简化您当前的方法。既然你说可以安全地假设这些是有效的公共域，我们可以定义一个这样的函数：
from pyspark.sql.functions import col, regexp_extract
def suffix(c):
return regexp_extract(c, "([^.]+\\.[^.]+$)", 1)
提取顶级域和一级子域：
domains_with_suffix = (domains
.withColumn("suffix", suffix("domains"))
.alias("domains"))
good_domains_with_suffix = (good_domains
.withColumn("suffix", suffix("gooddomains"))
.alias("good_domains"))
domains_with_suffix.show()
+--------------------+--------------------+
| domains| suffix|
+--------------------+--------------------+
|something.google.com| google.com|
|something.google....| ac.uk|
|something.good.co...| com.cy|
|something.good.co...| mal.org|
|something.bad.com.cy| com.cy|
| omgalsogood.com.cy| com.cy|
| good.com.cy| com.cy|
| sogood.example.com| example.com|
|notsoreal.googleecom|notsoreal.googleecom|
+--------------------+--------------------+
现在我们可以外连接了：
from pyspark.sql.functions import (
col, concat, lit, monotonically_increasing_id, sum as sum_
)
candidates = (domains_with_suffix
.join(
good_domains_with_suffix,
col("domains.suffix") == col("good_domains.suffix"),
"left"))
并过滤结果：
is_good_expr = (
col("good_domains.suffix").isNotNull() & # Match on suffix
(
# Exact match
(col("domains") == col("gooddomains")) |
# Subdomain match
col("domains").endswith(concat(lit("."), col("gooddomains")))
)
)
not_good_domains = (candidates
.groupBy("domains") # .groupBy("suffix", "domains") - see the discussion
.agg((sum_(is_good_expr.cast("integer")) > 0).alias("any_good"))
.filter(~col("any_good"))
.drop("any_good"))
not_good_domains.show(truncate=False)
+----------------------------------------+
|domains |
+----------------------------------------+
|omgalsogood.com.cy |
|notsoreal.googleecom |
|something.good.com.cy.mal.org |
|something.google.com.somethingelse.ac.uk|
|something.bad.com.cy |
+----------------------------------------+
这比Cartesian product required for direct join with LIKE 好，但无法满足蛮力，在最坏的情况下需要两次随机播放 - 一次用于join（如果good_domains 足够小到broadcasted，则可以跳过此操作)，另一个是group_by + agg。
不幸的是，Spark SQL 不允许自定义分区器对两者都只使用一个 shuffle（但是在 RDD API 中使用 composite key 是可能的）并且优化器还不够聪明，无法优化 join(_, "key1") 和 .groupBy("key1", _)。
如果您可以接受一些误报，则可以进行概率分析。首先让我们构建概率计数器（这里使用bounter 并得到toolz 的少量帮助）
from pyspark.sql.functions import concat_ws, reverse, split
from bounter import bounter
from toolz.curried import identity, partition_all
# This is only for testing on toy examples, in practice use more realistic value
size_mb = 20
chunk_size = 100
def reverse_domain(c):
return concat_ws(".", reverse(split(c, "\\.")))
def merge(acc, xs):
acc.update(xs)
return acc
counter = sc.broadcast((good_domains
.select(reverse_domain("gooddomains"))
.rdd.flatMap(identity)
# Chunk data into groups so we reduce the number of update calls
.mapPartitions(partition_all(chunk_size))
# Use tree aggregate to reduce pressure on the driver,
# when number of partitions is large*
# You can use depth parameter for further tuning
.treeAggregate(bounter(need_iteration=False, size_mb=size_mb), merge, merge)))
接下来定义一个像这样的用户定义函数函数
from pyspark.sql.functions import pandas_udf, PandasUDFType
from toolz import accumulate
def is_good_counter(counter):
def is_good_(x):
return any(
x in counter.value
for x in accumulate(lambda x, y: "{}.{}".format(x, y), x.split("."))
)
@pandas_udf("boolean", PandasUDFType.SCALAR)
def _(xs):
return xs.apply(is_good_)
return _
并过滤domains:
domains.filter(
~is_good_counter(counter)(reverse_domain("domains"))
).show(truncate=False)
+----------------------------------------+
|domains |
+----------------------------------------+
|something.google.com.somethingelse.ac.uk|
|something.good.com.cy.mal.org |
|something.bad.com.cy |
|omgalsogood.com.cy |
|notsoreal.googleecom |
+----------------------------------------+
在 Scala 中这可以通过 bloomFilter 完成
import org.apache.spark.sql.Column
import org.apache.spark.sql.functions._
import org.apache.spark.util.sketch.BloomFilter
def reverseDomain(c: Column) = concat_ws(".", reverse(split(c, "\\.")))
val checker = good_domains.stat.bloomFilter(
// Adjust values depending on the data
reverseDomain($"gooddomains"), 1000, 0.001
)
def isGood(checker: BloomFilter) = udf((s: String) =>
s.split('.').toStream.scanLeft("") {
case ("", x) => x
case (acc, x) => s"${acc}.${x}"
}.tail.exists(checker mightContain _))
domains.filter(!isGood(checker)(reverseDomain($"domains"))).show(false)
+----------------------------------------+
|domains |
+----------------------------------------+
|something.google.com.somethingelse.ac.uk|
|something.good.com.cy.mal.org |
|something.bad.com.cy |
|omgalsogood.com.cy |
|notsoreal.googleecom |
+----------------------------------------+
如果需要，shouldn't be hard to call such code from Python。
由于近似性质，这可能仍不能完全令人满意。如果您需要准确的结果，您可以尝试利用数据的冗余特性，例如使用trie（这里使用datrie 实现）。
如果good_domains 相对较小，您可以创建一个模型，方法与概率变体类似：
import string
import datrie
def seq_op(acc, x):
acc[x] = True
return acc
def comb_op(acc1, acc2):
acc1.update(acc2)
return acc1
trie = sc.broadcast((good_domains
.select(reverse_domain("gooddomains"))
.rdd.flatMap(identity)
# string.printable is a bit excessive if you need standard domain
# and not enough if you allow internationalized domain names.
# In the latter case you'll have to adjust the `alphabet`
# or use different implementation of trie.
.treeAggregate(datrie.Trie(string.printable), seq_op, comb_op)))
定义用户定义函数：
def is_good_trie(trie):
def is_good_(x):
if not x:
return False
else:
return any(
x == match or x[len(match)] == "."
for match in trie.value.iter_prefixes(x)
)
@pandas_udf("boolean", PandasUDFType.SCALAR)
def _(xs):
return xs.apply(is_good_)
return _
并将其应用于数据：
domains.filter(
~is_good_trie(trie)(reverse_domain("domains"))
).show(truncate=False)
+----------------------------------------+
|domains |
+----------------------------------------+
|something.google.com.somethingelse.ac.uk|
|something.good.com.cy.mal.org |
|something.bad.com.cy |
|omgalsogood.com.cy |
|notsoreal.googleecom |
+----------------------------------------+
这种特定方法的工作假设是所有good_domains 都可以压缩到一个单一的树中，但可以很容易地扩展以处理不满足此假设的情况。例如，您可以为每个顶级域或后缀（如在朴素解决方案中定义的那样）构建单个 trie
(good_domains
.select(suffix("gooddomains"), reverse_domain("gooddomains"))
.rdd
.aggregateByKey(datrie.Trie(string.printable), seq_op, comb_op))
然后，要么从序列化版本按需加载模型，要么使用RDD 操作。
这两种非原生方法可以根据数据、业务需求（如近似解的假阴性容忍度）和可用资源（驱动程序内存、执行程序内存、suffixes 的基数、访问分布式POSIX 兼容的分布式文件系统等）。在将这些应用于DataFrames 和RDDs（内存使用、通信和序列化开销）之间进行选择时，还需要考虑一些权衡。
* 见Understanding treeReduce() in Spark