def Compliance(self, instance, predicate, where=None):
result = self.AnalysisRunner.onData(self.df).addAnalyzer(Compliance(instance, predicate, where)).run()
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
result_json = AnalyzerContext.successMetricsAsJson(self.spark, result)
df_from_json = self.spark.read.json(self.sc.parallelize([result_json]))
self.assertEqual(df_from_json.select("value").collect(), result_df.select("value").collect())
return result_df.select("value").collect()
def UniqueValueRatio(self, columns, where=None):
result = self.AnalysisRunner.onData(self.df).addAnalyzer(UniqueValueRatio(columns, where)).run()
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
result_json = AnalyzerContext.successMetricsAsJson(self.spark, result)
df_from_json = self.spark.read.json(self.sc.parallelize([result_json]))
self.assertEqual(df_from_json.select("value").collect(), result_df.select("value").collect())
return result_df.select("value").collect()
def ApproxQuantiles(self, column, quantiles):
result = self.AnalysisRunner.onData(self.df).addAnalyzer(ApproxQuantiles(column, quantiles)).run()
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
result_json = AnalyzerContext.successMetricsAsJson(self.spark, result)
df_from_json = self.spark.read.json(self.sc.parallelize([result_json]))
self.assertEqual(df_from_json.select("value").collect(), result_df.select("value").collect())
return result_df.select("value").collect()
def test_PatternMatch(self):
result = (
self.AnalysisRunner.onData(self.df).addAnalyzer(PatternMatch(column="a", pattern_regex="ba(r|z)")).run()
)
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
result_json = AnalyzerContext.successMetricsAsJson(self.spark, result)
df_from_json = self.spark.read.json(self.sc.parallelize([result_json]))
self.assertEqual(df_from_json.select("value").collect(), result_df.select("value").collect())
self.assertEqual(result_df.select("value").collect(), [Row(value=0.0)])
def Histogram_maxBins(self, column, binningUdf=None, maxDetailBins: int = None, where: str = None):
result = (
self.AnalysisRunner.onData(self.df).addAnalyzer(Histogram(column, binningUdf, maxDetailBins, where)).run()
)
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
result_json = AnalyzerContext.successMetricsAsJson(self.spark, result)
df_from_json = self.spark.read.json(self.sc.parallelize([result_json]))
self.assertEqual(df_from_json.select("value").collect(), result_df.select("value").collect())
return result_df.select("value").collect()
def test_KLLSketch(self):
result = (
self.AnalysisRunner.onData(self.df).addAnalyzer(KLLSketch("b", KLLParameters(self.spark, 2, 0.64, 2))).run()
)
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
result_df.show()
return result_df.select("value").collect()
def __init__(
self,
dataframe: DataFrame,
spark: SparkSession,
profiling_config: DataLakeProfilerConfig,
report: DataLakeSourceReport,
file_path: str,
):
self.spark = spark
self.dataframe = dataframe
self.analyzer = AnalysisRunner(spark).onData(dataframe)
self.column_specs = []
self.row_count = dataframe.count()
self.profiling_config = profiling_config
self.file_path = file_path
self.columns_to_profile = []
self.ignored_columns = []
self.profile = DatasetProfileClass(timestampMillis=get_sys_time())
self.report = report
self.profile.rowCount = self.row_count
self.profile.columnCount = len(dataframe.columns)
column_types = {x.name: x.dataType for x in dataframe.schema.fields}
if self.profiling_config.profile_table_level_only:
return
# get column distinct counts
for column in dataframe.columns:
if not self.profiling_config.allow_deny_patterns.allowed(column):
self.ignored_columns.append(column)
continue
self.columns_to_profile.append(column)
# Normal CountDistinct is ridiculously slow
self.analyzer.addAnalyzer(ApproxCountDistinct(column))
if self.profiling_config.max_number_of_fields_to_profile is not None:
if (len(self.columns_to_profile) >
self.profiling_config.max_number_of_fields_to_profile):
columns_being_dropped = self.columns_to_profile[
self.profiling_config.max_number_of_fields_to_profile:]
self.columns_to_profile = self.columns_to_profile[:self.
profiling_config
.
max_number_of_fields_to_profile]
self.report.report_file_dropped(
f"The max_number_of_fields_to_profile={self.profiling_config.max_number_of_fields_to_profile} reached. Profile of columns {self.file_path}({', '.join(sorted(columns_being_dropped))})"
)
analysis_result = self.analyzer.run()
analysis_metrics = AnalyzerContext.successMetricsAsJson(
self.spark, analysis_result)
# reshape distinct counts into dictionary
column_distinct_counts = {
x["instance"]: int(x["value"])
for x in analysis_metrics if x["name"] == "ApproxCountDistinct"
}
select_numeric_null_counts = [
count(when(
isnan(c) | col(c).isNull(),
c,
)).alias(c) for c in self.columns_to_profile
if column_types[column] in [DoubleType, FloatType]
]
# PySpark doesn't support isnan() on non-float/double columns
select_nonnumeric_null_counts = [
count(when(
col(c).isNull(),
c,
)).alias(c) for c in self.columns_to_profile
if column_types[column] not in [DoubleType, FloatType]
]
null_counts = dataframe.select(select_numeric_null_counts +
select_nonnumeric_null_counts)
column_null_counts = null_counts.toPandas().T[0].to_dict()
column_null_fractions = {
c: column_null_counts[c] / self.row_count
for c in self.columns_to_profile
}
column_nonnull_counts = {
c: self.row_count - column_null_counts[c]
for c in self.columns_to_profile
}
column_unique_proportions = {
c:
(column_distinct_counts[c] /
column_nonnull_counts[c] if column_nonnull_counts[c] > 0 else 0)
for c in self.columns_to_profile
}
if self.profiling_config.include_field_sample_values:
# take sample and convert to Pandas DataFrame
if self.row_count < NUM_SAMPLE_ROWS:
# if row count is less than number to sample, just take all rows
rdd_sample = dataframe.rdd.take(self.row_count)
else:
rdd_sample = dataframe.rdd.takeSample(False,
NUM_SAMPLE_ROWS,
seed=0)
# init column specs with profiles
for column in self.columns_to_profile:
column_profile = DatasetFieldProfileClass(fieldPath=column)
column_spec = _SingleColumnSpec(column, column_profile)
column_profile.uniqueCount = column_distinct_counts.get(column)
column_profile.uniqueProportion = column_unique_proportions.get(
column)
column_profile.nullCount = column_null_counts.get(column)
column_profile.nullProportion = column_null_fractions.get(column)
if self.profiling_config.include_field_sample_values:
column_profile.sampleValues = [
str(x[column]) for x in rdd_sample
]
column_spec.type_ = column_types[column]
column_spec.cardinality = _convert_to_cardinality(
column_distinct_counts[column],
column_null_fractions[column],
)
self.column_specs.append(column_spec)
def ingest_table(self, full_path: str, relative_path: str,
is_aws: bool) -> Iterable[MetadataWorkUnit]:
table_name = self.get_table_name(relative_path, full_path)
# yield the table schema first
logger.debug(
f"Ingesting {full_path}: making table schemas {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
yield from self.get_table_schema(full_path, table_name, is_aws)
# If profiling is not enabled, skip the rest
if not self.source_config.profiling.enabled:
return
# read in the whole table with Spark for profiling
table = self.read_file_spark(full_path, is_aws)
# if table is not readable, skip
if table is None:
self.report.report_warning(
table_name,
f"unable to read table {table_name} from file {full_path}")
return
with PerfTimer() as timer:
# init PySpark analysis object
logger.debug(
f"Profiling {full_path}: reading file and computing nulls+uniqueness {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
table_profiler = _SingleTableProfiler(
table,
self.spark,
self.source_config.profiling,
self.report,
full_path,
)
logger.debug(
f"Profiling {full_path}: preparing profilers to run {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
# instead of computing each profile individually, we run them all in a single analyzer.run() call
# we use a single call because the analyzer optimizes the number of calls to the underlying profiler
# since multiple profiles reuse computations, this saves a lot of time
table_profiler.prepare_table_profiles()
# compute the profiles
logger.debug(
f"Profiling {full_path}: computing profiles {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
analysis_result = table_profiler.analyzer.run()
analysis_metrics = AnalyzerContext.successMetricsAsDataFrame(
self.spark, analysis_result)
logger.debug(
f"Profiling {full_path}: extracting profiles {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
table_profiler.extract_table_profiles(analysis_metrics)
time_taken = timer.elapsed_seconds()
logger.info(
f"Finished profiling {full_path}; took {time_taken:.3f} seconds"
)
self.profiling_times_taken.append(time_taken)
mcp = MetadataChangeProposalWrapper(
entityType="dataset",
entityUrn=make_dataset_urn(self.source_config.platform, table_name,
self.source_config.env),
changeType=ChangeTypeClass.UPSERT,
aspectName="datasetProfile",
aspect=table_profiler.profile,
)
wu = MetadataWorkUnit(
id=f"profile-{self.source_config.platform}-{full_path}", mcp=mcp)
self.report.report_workunit(wu)
yield wu
def get_table_profile(self, table_data: TableData,
dataset_urn: str) -> Iterable[MetadataWorkUnit]:
# read in the whole table with Spark for profiling
table = None
try:
table = self.read_file_spark(
table_data.table_path,
os.path.splitext(table_data.full_path)[1])
except Exception as e:
logger.error(e)
# if table is not readable, skip
if table is None:
self.report.report_warning(
table_data.display_name,
f"unable to read table {table_data.display_name} from file {table_data.full_path}",
)
return
with PerfTimer() as timer:
# init PySpark analysis object
logger.debug(
f"Profiling {table_data.full_path}: reading file and computing nulls+uniqueness {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
table_profiler = _SingleTableProfiler(
table,
self.spark,
self.source_config.profiling,
self.report,
table_data.full_path,
)
logger.debug(
f"Profiling {table_data.full_path}: preparing profilers to run {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
# instead of computing each profile individually, we run them all in a single analyzer.run() call
# we use a single call because the analyzer optimizes the number of calls to the underlying profiler
# since multiple profiles reuse computations, this saves a lot of time
table_profiler.prepare_table_profiles()
# compute the profiles
logger.debug(
f"Profiling {table_data.full_path}: computing profiles {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
analysis_result = table_profiler.analyzer.run()
analysis_metrics = AnalyzerContext.successMetricsAsDataFrame(
self.spark, analysis_result)
logger.debug(
f"Profiling {table_data.full_path}: extracting profiles {datetime.now().strftime('%d/%m/%Y %H:%M:%S')}"
)
table_profiler.extract_table_profiles(analysis_metrics)
time_taken = timer.elapsed_seconds()
logger.info(
f"Finished profiling {table_data.full_path}; took {time_taken:.3f} seconds"
)
self.profiling_times_taken.append(time_taken)
mcp = MetadataChangeProposalWrapper(
entityType="dataset",
entityUrn=dataset_urn,
changeType=ChangeTypeClass.UPSERT,
aspectName="datasetProfile",
aspect=table_profiler.profile,
)
wu = MetadataWorkUnit(
id=f"profile-{self.source_config.platform}-{table_data.table_path}",
mcp=mcp)
self.report.report_workunit(wu)
yield wu
def test_Size(self):
result = self.AnalysisRunner.onData(self.df).addAnalyzer(Size()).run()
# result_df = result.select('value').collect()
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
result_df_row = result_df.select("value").collect()
self.assertEqual(result_df_row, [Row(value=3.0)])
def Correlation(self, column1, column2, where=None):
result = self.AnalysisRunner.onData(self.df).addAnalyzer(Correlation(column1, column2, where)).run()
result_df = AnalyzerContext.successMetricsAsDataFrame(self.spark, result)
AnalyzerContext.successMetricsAsJson(self.spark, result)
return result_df.select("value").collect()
