def inner_join_streaming():
s_env = StreamExecutionEnvironment.get_execution_environment()
s_env.set_parallelism(1)
st_env = StreamTableEnvironment.create(s_env)
result_file = "/tmp/table_inner_join_streaming.csv"
if os.path.exists(result_file):
os.remove(result_file)
left = st_env.from_elements([(1, "1a", "1laa"), (2, "2a", "2aa"),
(3, None, "3aa"), (2, "4b", "4bb"),
(5, "5a", "5aa")],
["a", "b", "c"]).select("a, b, c")
right = st_env.from_elements([(1, "1b", "1bb"), (2, None, "2bb"),
(1, "3b", "3bb"), (4, "4b", "4bb")],
["d", "e", "f"]).select("d, e, f")
st_env.register_table_sink(
"result",
CsvTableSink(
["a", "b", "c"],
[DataTypes.BIGINT(),
DataTypes.STRING(),
DataTypes.STRING()], result_file))
result = left.join(right).where("a = d").select("a, b, e")
result.insert_into("result")
st_env.execute("inner join streaming")
def add_train_chief_alone_table():
stream_env = StreamExecutionEnvironment.get_execution_environment()
table_env = StreamTableEnvironment.create(stream_env)
statement_set = table_env.create_statement_set()
work_num = 2
ps_num = 1
python_file = os.getcwd() + "/../../src/test/python/add.py"
func = "map_func"
prop = {}
prop[TFCONSTANS.TF_IS_CHIEF_ALONE] = "true"
prop[MLCONSTANTS.PYTHON_VERSION] = "3.7"
env_path = None
input_tb = None
output_schema = None
tf_config = TFConfig(work_num, ps_num, prop, python_file, func,
env_path)
train(stream_env, table_env, statement_set, input_tb, tf_config,
output_schema)
# inference(stream_env, table_env, statement_set, input_tb, tf_config, output_schema)
job_client = statement_set.execute().get_job_client()
if job_client is not None:
job_client.get_job_execution_result(
user_class_loader=None).result()
def test_execute(self):
tmp_dir = tempfile.gettempdir()
field_names = ['a', 'b', 'c']
field_types = [
DataTypes.BIGINT(),
DataTypes.STRING(),
DataTypes.STRING()
]
t_env = StreamTableEnvironment.create(self.env)
t_env.register_table_sink(
'Results',
CsvTableSink(
field_names, field_types,
os.path.join('{}/{}.csv'.format(tmp_dir, round(time.time())))))
t_env.insert_into(
'Results',
t_env.from_elements([(1, 'Hi', 'Hello')], ['a', 'b', 'c']))
execution_result = t_env.execute('test_stream_execute')
self.assertIsNotNone(execution_result.get_job_id())
self.assertIsNotNone(execution_result.get_net_runtime())
self.assertEqual(len(execution_result.get_all_accumulator_results()),
0)
self.assertIsNone(
execution_result.get_accumulator_result('accumulator'))
self.assertIsNotNone(str(execution_result))
def __init__(self):
# self.feature_extractor = DemoFeatureExtractor()
self.settings = EnvironmentSettings.new_instance().in_streaming_mode(
).use_blink_planner().build()
self.env = StreamExecutionEnvironment.get_execution_environment()
self.env.set_parallelism(1)
self.table_env = StreamTableEnvironment.create(
self.env, environment_settings=self.settings)
self.table_env.get_config().get_configuration().set_boolean(
"python.fn-execution.memory.managed", True)
self.table_env.get_config().get_configuration().set_string(
"python.fn-execution.buffer.memory.size", "1024mb")
self.table_env.get_config().get_configuration().set_string(
"parallelism.default", "3")
self.table_env.get_config().get_configuration().set_string(
"python.fn-execution.bundle.size", "5000")
self.table_env.get_config().get_configuration().set_string(
"restart-strategy", "fixed-delay")
self.table_env.get_config().get_configuration().set_string(
"restart-strategy.fixed-delay.attempts", "3")
self.table_env.get_config().get_configuration().set_string(
"restart-strategy.fixed-delay.delay", "30s")
source_table = open('source.sql', 'r').read()
sink_table = open('sink.sql', 'r').read()
self.table_env.execute_sql(source_table)
self.table_env.execute_sql(sink_table)
def select_streaming():
s_env = StreamExecutionEnvironment.get_execution_environment()
s_env.set_parallelism(1)
st_env = StreamTableEnvironment.create(s_env)
source_file = os.getcwd() + "/../resources/table_orders.csv"
result_file = "/tmp/table_select_streaming.csv"
if os.path.exists(result_file):
os.remove(result_file)
st_env.register_table_source(
"Orders",
CsvTableSource(source_file, ["a", "b", "c", "rowtime"], [
DataTypes.STRING(),
DataTypes.INT(),
DataTypes.INT(),
DataTypes.TIMESTAMP()
]))
st_env.register_table_sink(
"result",
CsvTableSink(["a", "c"],
[DataTypes.STRING(), DataTypes.INT()], result_file))
orders = st_env.scan("Orders")
result = orders.select("a, b")
result.insert_into("result")
st_env.execute("select streaming")
def __init__(self,
parallelism: int = 1,
checkpoint_interval: Optional[int] = None,
state_ttl: Optional[int] = None) -> None:
# setting env
env = StreamExecutionEnvironment.get_execution_environment()
env.set_parallelism(parallelism)
if checkpoint_interval:
env.set_state_backend(
RocksDBStateBackend(self.checkpoints_path,
enable_incremental_checkpointing=True))
env.enable_checkpointing(checkpoint_interval * 1000)
t_config = TableConfig()
if state_ttl:
t_config.set_idle_state_retention_time(
timedelta(seconds=state_ttl),
timedelta(seconds=state_ttl + 300))
table_env = StreamTableEnvironment.create(env, table_config=t_config)
table_env.get_config().get_configuration().set_string(
"pipeline.jars", self.flink_sql_connector_kafka_jar)
# set up table
for ddl in self.tables:
table_env.execute_sql(ddl)
self.env = env
self.table_env = table_env
def main():
env = StreamExecutionEnvironment.get_execution_environment()
env.set_parallelism(1)
environment_settings = EnvironmentSettings.new_instance().use_blink_planner().build()
t_env = StreamTableEnvironment.create(env, environment_settings=environment_settings)
t_env.get_config().get_configuration().set_integer("python.fn-execution.bundle.size", 300000)
t_env.get_config().get_configuration().set_integer("python.fn-execution.bundle.time", 1000)
t_env.get_config().get_configuration().set_boolean("pipeline.object-reuse", True)
t_env.register_table_sink(
"sink",
PrintTableSink(
["id"],
[DataTypes.INT(False)]))
@udf(input_types=[DataTypes.INT(False)], result_type=DataTypes.INT(False))
def inc(x):
return x + 1
t_env.register_function("inc", inc)
t_env.register_java_function("java_inc", "com.alibaba.flink.function.JavaInc")
num_rows = 100000000
t_env.from_table_source(RangeTableSource(1, num_rows, 1)).alias("id") \
.select("inc(id)") \
.insert_into("sink")
beg_time = time.time()
t_env.execute("Python UDF")
print("PyFlink Python UDF inc() consume time: " + str(time.time() - beg_time))
def test_table_environment_with_blink_planner(self):
self.env.set_parallelism(1)
t_env = StreamTableEnvironment.create(
self.env,
environment_settings=EnvironmentSettings.new_instance(
).use_blink_planner().build())
source_path = os.path.join(self.tempdir + '/streaming.csv')
sink_path = os.path.join(self.tempdir + '/result.csv')
field_names = ["a", "b", "c"]
field_types = [DataTypes.INT(), DataTypes.STRING(), DataTypes.STRING()]
data = [(1, 'hi', 'hello'), (2, 'hello', 'hello')]
csv_source = self.prepare_csv_source(source_path, data, field_types,
field_names)
t_env.register_table_source("source", csv_source)
t_env.register_table_sink(
"sink", CsvTableSink(field_names, field_types, sink_path))
source = t_env.scan("source")
result = source.alias("a, b, c").select("1 + a, b, c")
result.insert_into("sink")
t_env.execute("blink_test")
results = []
with open(sink_path, 'r') as f:
results.append(f.readline())
results.append(f.readline())
self.assert_equals(results, ['2,hi,hello\n', '3,hello,hello\n'])
def main_flink():
#之前的步骤是拿文件然后预处理然后写文件到input这个文件中
env = StreamExecutionEnvironment.get_execution_environment()
parr_num = 4
env.set_parallelism(parr_num)
t_env = StreamTableEnvironment.create(env)
@udf(input_types=DataTypes.STRING(), result_type=DataTypes.STRING())
def cut_extract(string):
return cut_posseg.cut_extract(string)
t_env.register_function("cut_extract",cut_extract)
#这里是建表然后从input拿，问题1：有没有办法从自定义的list来当做输入来节省IO开销呢，比如我输入[文本A，文本B]这样的list作为输入
t_env.connect(FileSystem().path('/home/sjtu/input')) \
.with_format(OldCsv()
.field('text', DataTypes.STRING())) \
.with_schema(Schema()
.field('text', DataTypes.STRING())) \
.create_temporary_table('mySource')
t_env.connect(FileSystem().path('/home/sjtu/output')) \
.with_format(OldCsv()
.field('result', DataTypes.STRING())) \
.with_schema(Schema()
.field('result', DataTypes.STRING())) \
.create_temporary_table('mySink')
t_env.from_path('mySource')\
.select("cut_extract(text)")\
.insert_into('mySink')
#问题2：这里我是将结果的表写了文件，但实际上我还需要在这个代码中继续对这些处理完的数据进行处理，也没有办法直接将上述的mySink表直接
#作为内存数据取出来而不是从硬盘再读入呢
t_env.execute("tutorial_job")
def init_env(self, **kwargs):
env = StreamExecutionEnvironment.get_execution_environment()
self.st_env = StreamTableEnvironment.create(
env,
environment_settings = EnvironmentSettings.new_instance().use_blink_planner().build()
)
return
def train(stream_env=None,
table_env=None,
statement_set=None,
input_table=None,
tf_config=None,
output_schema=None):
"""
Run TF train for flink table api.
:param stream_env: The StreamExecutionEnvironment. If it's None, this method will create one and execute the job
at the end. Otherwise, caller is responsible to trigger the job execution
:param table_env: The Flink TableEnvironment.
:param statement_set: The StatementSet created by the given TableEnvironment.
:param input_table: The input Table.
:param tf_config: Configurations for the TF program.
:param output_schema: The TableSchema for the output Table. If it's null, a dummy sink will be connected.
:return: output Table. Otherwise, caller is responsible to add sink to the output Table before executing the graph.
"""
if stream_env is None:
stream_env = StreamExecutionEnvironment.get_execution_environment()
if table_env is None:
table_env = StreamTableEnvironment.create(stream_env)
if statement_set is None:
statement_set = table_env.create_statement_set()
if input_table is not None:
input_table = input_table._j_table
if output_schema is not None:
output_schema = output_schema._j_table_schema
output_table = get_gateway(
).jvm.org.flinkextended.flink.ml.tensorflow.client.TFUtils.train(
stream_env._j_stream_execution_environment,
table_env._j_tenv, statement_set._j_statement_set, input_table,
tf_config.java_config(), output_schema)
return Table(output_table, table_env)
def test_custom_env(self):
import pyflink
from pyflink.dataset import ExecutionEnvironment
from pyflink.datastream import StreamExecutionEnvironment
benv = ExecutionEnvironment.get_execution_environment()
senv = StreamExecutionEnvironment.get_execution_environment()
from pyflink.table import BatchTableEnvironment
from pyflink.table import StreamTableEnvironment
btenv = BatchTableEnvironment.create(benv)
stenv = StreamTableEnvironment.create(senv)
mlenv = useCustomEnv(pyflink.java_gateway.get_gateway(),
benv, btenv, senv, stenv)
t = mlenv.btenv.from_elements([(1, 2), (2, 5), (3, 1)], ['a', 'b'])
source = TableSourceBatchOp(t)
source.print()
t = mlenv.stenv.from_elements([(1, 2), (2, 5), (3, 1)], ['a', 'b'])
source = TableSourceStreamOp(t)
source.print()
StreamOperator.execute()
from pyalink.alink import env
env._in_custom_env = False
resetEnv()
def full_outer_join_streaming():
s_env = StreamExecutionEnvironment.get_execution_environment()
s_env.set_parallelism(1)
# use blink table planner
st_env = StreamTableEnvironment.create(
s_env,
environment_settings=EnvironmentSettings.new_instance(
).in_streaming_mode().use_blink_planner().build())
# use flink table planner
# st_env = StreamTableEnvironment.create(s_env)
left = st_env.from_elements([(1, "1a", "1laa"), (2, "2a", "2aa"),
(3, None, "3aa"), (2, "4b", "4bb"),
(5, "5a", "5aa")],
["a", "b", "c"]).select("a, b, c")
right = st_env.from_elements([(1, "1b", "1bb"), (2, None, "2bb"),
(1, "3b", "3bb"), (4, "4b", "4bb")],
["d", "e", "f"]).select("d, e, f")
result = left.full_outer_join(right, "a = d").select("a, b, e")
# use custom retract sink connector
sink = TestRetractSink(
["a", "b", "c"],
[DataTypes.BIGINT(),
DataTypes.STRING(),
DataTypes.STRING()])
st_env.register_table_sink("sink", sink)
result.insert_into("sink")
st_env.execute("full outer join streaming")
def input_output_table():
stream_env = StreamExecutionEnvironment.get_execution_environment()
table_env = StreamTableEnvironment.create(stream_env)
statement_set = table_env.create_statement_set()
work_num = 2
ps_num = 1
python_file = os.getcwd() + "/../../src/test/python/input_output.py"
prop = {}
func = "map_func"
env_path = None
prop[
MLCONSTANTS.
ENCODING_CLASS] = "org.flinkextended.flink.ml.operator.coding.RowCSVCoding"
prop[
MLCONSTANTS.
DECODING_CLASS] = "org.flinkextended.flink.ml.operator.coding.RowCSVCoding"
inputSb = "INT_32" + "," + "INT_64" + "," + "FLOAT_32" + "," + "FLOAT_64" + "," + "STRING"
prop["sys:csv_encode_types"] = inputSb
prop["sys:csv_decode_types"] = inputSb
prop[MLCONSTANTS.PYTHON_VERSION] = "3.7"
source_file = os.getcwd() + "/../../src/test/resources/input.csv"
sink_file = os.getcwd() + "/../../src/test/resources/output.csv"
table_source = CsvTableSource(source_file, ["a", "b", "c", "d", "e"], [
DataTypes.INT(),
DataTypes.BIGINT(),
DataTypes.FLOAT(),
DataTypes.DOUBLE(),
DataTypes.STRING()
])
table_env.register_table_source("source", table_source)
input_tb = table_env.from_path("source")
output_schema = TableSchema(["a", "b", "c", "d", "e"], [
DataTypes.INT(),
DataTypes.BIGINT(),
DataTypes.FLOAT(),
DataTypes.DOUBLE(),
DataTypes.STRING()
])
sink = CsvTableSink(["a", "b", "c", "d", "e"], [
DataTypes.INT(),
DataTypes.BIGINT(),
DataTypes.FLOAT(),
DataTypes.DOUBLE(),
DataTypes.STRING()
],
sink_file,
write_mode=WriteMode.OVERWRITE)
table_env.register_table_sink("table_row_sink", sink)
tf_config = TFConfig(work_num, ps_num, prop, python_file, func,
env_path)
output_table = train(stream_env, table_env, statement_set, input_tb,
tf_config, output_schema)
# output_table = inference(stream_env, table_env, statement_set, input_tb, tf_config, output_schema)
statement_set.add_insert("table_row_sink", output_table)
job_client = statement_set.execute().get_job_client()
if job_client is not None:
job_client.get_job_execution_result(
user_class_loader=None).result()
def filter_streaming():
s_env = StreamExecutionEnvironment.get_execution_environment()
s_env.set_parallelism(1)
# use blink table planner
st_env = StreamTableEnvironment.create(
s_env,
environment_settings=EnvironmentSettings.new_instance(
).in_streaming_mode().use_blink_planner().build())
# use flink table planner
# st_env = StreamTableEnvironment.create(s_env)
source_file = os.getcwd() + "/../resources/table_orders.csv"
result_file = "/tmp/table_filter_streaming.csv"
if os.path.exists(result_file):
os.remove(result_file)
st_env.register_table_source(
"Orders",
CsvTableSource(source_file, ["a", "b", "c", "rowtime"], [
DataTypes.STRING(),
DataTypes.INT(),
DataTypes.INT(),
DataTypes.TIMESTAMP()
]))
st_env.register_table_sink(
"result",
CsvTableSink(["a", "b", "c", "rowtime"], [
DataTypes.STRING(),
DataTypes.INT(),
DataTypes.INT(),
DataTypes.TIMESTAMP()
], result_file))
orders = st_env.scan("Orders")
result = orders.filter("b % 2 === 0")
result.insert_into("result")
st_env.execute("filter streaming")
def log_processing():
env = StreamExecutionEnvironment.get_execution_environment()
env_settings = EnvironmentSettings.Builder().use_blink_planner().build()
t_env = StreamTableEnvironment.create(stream_execution_environment=env,
environment_settings=env_settings)
t_env.get_config().get_configuration().set_boolean(
"python.fn-execution.memory.managed", True)
source_ddl = """
CREATE TABLE payment_msg(
createTime VARCHAR,
orderId BIGINT,
payAmount DOUBLE,
payPlatform INT,
provinceId INT
) WITH (
'connector.type' = 'kafka',
'connector.version' = 'universal',
'connector.topic' = 'payment_msg',
'connector.properties.bootstrap.servers' = 'kafka:9092',
'connector.properties.group.id' = 'test_3',
'connector.startup-mode' = 'latest-offset',
'format.type' = 'json'
)
"""
es_sink_ddl = """
CREATE TABLE es_sink (
province VARCHAR PRIMARY KEY,
pay_amount DOUBLE
) with (
'connector.type' = 'elasticsearch',
'connector.version' = '7',
'connector.hosts' = 'http://elasticsearch:9200',
'connector.index' = 'platform_pay_amount_1',
'connector.document-type' = 'payment',
'update-mode' = 'upsert',
'connector.flush-on-checkpoint' = 'true',
'connector.key-delimiter' = '$',
'connector.key-null-literal' = 'n/a',
'connector.bulk-flush.max-size' = '42mb',
'connector.bulk-flush.max-actions' = '32',
'connector.bulk-flush.interval' = '1000',
'connector.bulk-flush.backoff.delay' = '1000',
'format.type' = 'json'
)
"""
t_env.sql_update(source_ddl)
t_env.sql_update(es_sink_ddl)
t_env.register_function('province_id_to_name', province_id_to_name)
t_env.from_path("payment_msg") \
.select("province_id_to_name(provinceId) as province, payAmount") \
.group_by("province") \
.select("province, sum(payAmount) as pay_amount") \
.insert_into("es_sink")
t_env.execute("payment_demo")
def setUp(self):
super(PyFlinkBlinkStreamTableTestCase, self).setUp()
self.env = StreamExecutionEnvironment.get_execution_environment()
self.env.set_parallelism(2)
self.t_env = StreamTableEnvironment.create(
self.env,
environment_settings=EnvironmentSettings.new_instance(
).in_streaming_mode().use_blink_planner().build())
def create_env(
self) -> (ExecutionEnvironment, TableEnvironment, StatementSet):
exec_env = StreamExecutionEnvironment.get_execution_environment()
exec_env.set_parallelism(1)
t_config = TableConfig()
t_env = StreamTableEnvironment.create(exec_env, t_config)
t_env.get_config().get_configuration().set_string(
"taskmanager.memory.task.off-heap.size", '80m')
statement_set = t_env.create_statement_set()
return exec_env, t_env, statement_set
def test_create_table_environment_with_blink_planner(self):
t_env = StreamTableEnvironment.create(
self.env,
environment_settings=EnvironmentSettings.new_instance().use_blink_planner().build())
planner = t_env._j_tenv.getPlanner()
self.assertEqual(
planner.getClass().getName(),
"org.apache.flink.table.planner.delegation.StreamPlanner")
def group_by_agg_streaming():
s_env = StreamExecutionEnvironment.get_execution_environment()
s_env.set_parallelism(1)
# use blink table planner
st_env = StreamTableEnvironment.create(
s_env,
environment_settings=EnvironmentSettings.new_instance(
).in_streaming_mode().use_blink_planner().build())
# use flink table planner
# st_env = StreamTableEnvironment.create(s_env)
source_file = os.getcwd() + "/../resources/table_orders.csv"
st_env.register_table_source(
"Orders",
CsvTableSource(source_file, ["a", "b", "c", "rowtime"], [
DataTypes.STRING(),
DataTypes.INT(),
DataTypes.INT(),
DataTypes.TIMESTAMP()
]))
st_env.connect(
Elasticsearch()
.version("6")
.host("localhost", 9200, "http")
.index("group_by_agg_streaming")
.document_type('pyflink')
.key_delimiter("_")
.key_null_literal("null")
.failure_handler_ignore()
.disable_flush_on_checkpoint()
.bulk_flush_max_actions(2)
.bulk_flush_max_size("1 mb")
.bulk_flush_interval(5000)
) \
.with_schema(
Schema()
.field("a", DataTypes.STRING())
.field("b", DataTypes.STRING())
) \
.with_format(
Json()
.derive_schema()
) \
.in_upsert_mode() \
.register_table_sink("result")
orders = st_env.scan("Orders")
groub_by_table = orders.group_by("a").select("a, b.sum as d")
# Because the schema of index user in elasticsearch is
# {"a":{"type":"text","fields":{"keyword":{"type":"keyword","ignore_above":256}}},
# "b":{"type":"text","fields":{"keyword":{"type":"keyword","ignore_above":256}}}
# so we need to cast the type in our demo.
st_env.register_table("group_table", groub_by_table)
result = st_env.sql_query("SELECT a, CAST(d AS VARCHAR) from group_table")
result.insert_into("result")
st_env.execute("group by agg streaming")
def test_construct_with_stream_env(self):
stream_execution_environment = StreamExecutionEnvironment.get_execution_environment()
stream_table_environment = StreamTableEnvironment.create(stream_execution_environment)
ml_environment = MLEnvironment(
stream_exe_env=stream_execution_environment,
stream_tab_env=stream_table_environment)
self.assertEqual(
ml_environment.get_stream_execution_environment(),
stream_execution_environment)
self.assertEqual(ml_environment.get_stream_table_environment(), stream_table_environment)
def get_stream_table_environment(self) -> StreamTableEnvironment:
"""
Get the StreamTableEnvironment. If the StreamTableEnvironment has not been set,
it initial the StreamTableEnvironment with default Configuration.
:return: the StreamTableEnvironment.
"""
if self._stream_tab_env is None:
self._stream_tab_env = StreamTableEnvironment.create(
StreamExecutionEnvironment.get_execution_environment())
return self._stream_tab_env
def setUp(self):
self.env = StreamExecutionEnvironment.get_execution_environment()
self._load_dependency_jars()
config = Configuration(
j_configuration=get_j_env_configuration(self.env._j_stream_execution_environment))
config.set_boolean("execution.checkpointing.checkpoints-after-tasks-finish.enabled", True)
self.env.set_parallelism(4)
self.env.enable_checkpointing(100)
self.env.set_restart_strategy(RestartStrategies.no_restart())
self.t_env = StreamTableEnvironment.create(self.env)
self.temp_dir = tempfile.mkdtemp()
def setUp(self) -> None:
from pyflink.datastream import StreamExecutionEnvironment
super(DataStreamConversionTestCases, self).setUp()
config = Configuration()
config.set_string("akka.ask.timeout", "20 s")
self.env = StreamExecutionEnvironment.get_execution_environment(config)
self.t_env = StreamTableEnvironment.create(self.env)
self.env.set_parallelism(2)
self.t_env.get_config().set("python.fn-execution.bundle.size", "1")
self.test_sink = DataStreamTestSinkFunction()
def pandas_udaf():
env = StreamExecutionEnvironment.get_execution_environment()
env.set_parallelism(1)
t_env = StreamTableEnvironment.create(stream_execution_environment=env)
# define the source with watermark definition
ds = env.from_collection(
collection=[
(Instant.of_epoch_milli(1000), 'Alice', 110.1),
(Instant.of_epoch_milli(4000), 'Bob', 30.2),
(Instant.of_epoch_milli(3000), 'Alice', 20.0),
(Instant.of_epoch_milli(2000), 'Bob', 53.1),
(Instant.of_epoch_milli(5000), 'Alice', 13.1),
(Instant.of_epoch_milli(3000), 'Bob', 3.1),
(Instant.of_epoch_milli(7000), 'Bob', 16.1),
(Instant.of_epoch_milli(10000), 'Alice', 20.1)
],
type_info=Types.ROW([Types.INSTANT(), Types.STRING(), Types.FLOAT()]))
table = t_env.from_data_stream(
ds,
Schema.new_builder()
.column_by_expression("ts", "CAST(f0 AS TIMESTAMP_LTZ(3))")
.column("f1", DataTypes.STRING())
.column("f2", DataTypes.FLOAT())
.watermark("ts", "ts - INTERVAL '3' SECOND")
.build()
).alias("ts, name, price")
# define the sink
t_env.create_temporary_table(
'sink',
TableDescriptor.for_connector('print')
.schema(Schema.new_builder()
.column('name', DataTypes.STRING())
.column('total_price', DataTypes.FLOAT())
.column('w_start', DataTypes.TIMESTAMP_LTZ())
.column('w_end', DataTypes.TIMESTAMP_LTZ())
.build())
.build())
@udaf(result_type=DataTypes.FLOAT(), func_type="pandas")
def mean_udaf(v):
return v.mean()
# define the tumble window operation
table = table.window(Tumble.over(lit(5).seconds).on(col("ts")).alias("w")) \
.group_by(table.name, col('w')) \
.select(table.name, mean_udaf(table.price), col("w").start, col("w").end)
# submit for execution
table.execute_insert('sink') \
.wait()
def create_table_env(self):
stream_env = StreamExecutionEnvironment.get_execution_environment()
stream_env.set_parallelism(1)
t_env = StreamTableEnvironment.create(
stream_env,
environment_settings=EnvironmentSettings.new_instance(
).in_streaming_mode().use_blink_planner().build())
statement_set = t_env.create_statement_set()
t_env.get_config().set_python_executable('/usr/bin/python3')
t_env.get_config().get_configuration().set_boolean(
"python.fn-execution.memory.managed", True)
return stream_env, t_env, statement_set
def hello_world():
"""
从随机Source读取数据，然后直接利用PrintSink输出。
"""
settings = EnvironmentSettings.new_instance().in_streaming_mode(
).use_blink_planner().build()
env = StreamExecutionEnvironment.get_execution_environment()
t_env = StreamTableEnvironment.create(stream_execution_environment=env,
environment_settings=settings)
t_env.get_config().get_configuration().set_boolean(
"python.fn-execution.memory.managed", True)
source_ddl = """
CREATE TABLE random_source (
f_sequence INT,
f_random INT,
f_random_str STRING
) WITH (
'connector' = 'datagen',
'rows-per-second'='5',
'fields.f_sequence.kind'='sequence',
'fields.f_sequence.start'='1',
'fields.f_sequence.end'='1000',
'fields.f_random.min'='1',
'fields.f_random.max'='1000',
'fields.f_random_str.length'='10'
)
"""
sink_ddl = """
CREATE TABLE print_sink (
f_sequence INT,
f_random INT,
f_random_str STRING
) WITH (
'connector' = 'print'
)
"""
# 注册source和sink
t_env.execute_sql(source_ddl)
t_env.execute_sql(sink_ddl)
# 注册 UDF
t_env.register_function('pass_by', pass_by)
# 数据提取
tab = t_env.from_path("random_source")
# 这里我们暂时先使用 标注了 deprecated 的API, 因为新的异步提交测试有待改进...
tab.select("f_sequence, f_random, pass_by(f_random_str) ").insert_into(
"print_sink")
# 执行作业
t_env.execute("Flink Hello World")
def main():
env = StreamExecutionEnvironment.get_execution_environment()
env.add_jars("file:///app/src/kafka-clients-2.8.0.jar")
env.add_jars("file:///app/src/flink-connector-kafka_2.12-1.12.3.jar")
env.set_parallelism(1)
env.set_stream_time_characteristic(TimeCharacteristic.EventTime)
env.enable_checkpointing(60000, CheckpointingMode.EXACTLY_ONCE)
config = env.get_checkpoint_config()
config.enable_externalized_checkpoints(
ExternalizedCheckpointCleanup.DELETE_ON_CANCELLATION)
st_env = StreamTableEnvironment.create(
env,
environment_settings=EnvironmentSettings.new_instance(
).in_streaming_mode().use_blink_planner().build())
print("register kafka source")
register_kafka_source(st_env)
print("register transaction sinks")
register_transactions_sink_into_csv(st_env)
st_env.from_path("source_tbl") \
.window(Slide.over(lit(1).minute).every(lit(5).seconds).on("ts").alias("w")) \
.group_by(col("w")) \
.select("""count(message) as total,
w.end as end_time
""") \
.insert_into("total_sink")
st_env.from_path("source_tbl") \
.where("message = 'dolorem'") \
.window(Slide.over(lit(1).minute).every(lit(5).seconds).on("ts").alias("w")) \
.group_by(col("w")) \
.select("""
count(message) as total,
w.end as end_time
""") \
.insert_into("grep_sink")
st_env.from_path("source_tbl") \
.window(Slide.over(lit(1).minute).every(lit(5).seconds).on("ts").alias("w")) \
.group_by(col("w"), col("message")) \
.select("""
count(message) as total,
message,
w.end as end_time
""") \
.insert_into("topk_sink")
st_env.execute("app")
def test_add_python_file(self):
import uuid
env = self.env
python_file_dir = os.path.join(self.tempdir,
"python_file_dir_" + str(uuid.uuid4()))
os.mkdir(python_file_dir)
python_file_path = os.path.join(python_file_dir, "test_dep1.py")
with open(python_file_path, 'w') as f:
f.write("def add_two(a):\n return a + 2")
def plus_two_map(value):
from test_dep1 import add_two
return add_two(value)
get_j_env_configuration(env._j_stream_execution_environment).\
setString("taskmanager.numberOfTaskSlots", "10")
env.add_python_file(python_file_path)
ds = env.from_collection([1, 2, 3, 4, 5])
ds = ds.map(plus_two_map, Types.LONG()) \
.slot_sharing_group("data_stream") \
.map(lambda i: i, Types.LONG()) \
.slot_sharing_group("table")
python_file_path = os.path.join(python_file_dir, "test_dep2.py")
with open(python_file_path, 'w') as f:
f.write("def add_three(a):\n return a + 3")
def plus_three(value):
from test_dep2 import add_three
return add_three(value)
t_env = StreamTableEnvironment.create(
stream_execution_environment=env,
environment_settings=EnvironmentSettings.in_streaming_mode())
env.add_python_file(python_file_path)
from pyflink.table.udf import udf
from pyflink.table.expressions import col
add_three = udf(plus_three, result_type=DataTypes.BIGINT())
tab = t_env.from_data_stream(ds, col('a')) \
.select(add_three(col('a')))
t_env.to_append_stream(tab, Types.ROW([Types.LONG()])) \
.map(lambda i: i[0]) \
.add_sink(self.test_sink)
env.execute("test add_python_file")
result = self.test_sink.get_results(True)
expected = ['6', '7', '8', '9', '10']
result.sort()
expected.sort()
self.assertEqual(expected, result)
def tumble_window_demo():
env = StreamExecutionEnvironment.get_execution_environment()
env.set_parallelism(1)
t_env = StreamTableEnvironment.create(stream_execution_environment=env)
# define the source with watermark definition
ds = env.from_collection(
collection=[
(Instant.of_epoch_milli(1000), 'Alice', 110.1),
(Instant.of_epoch_milli(4000), 'Bob', 30.2),
(Instant.of_epoch_milli(3000), 'Alice', 20.0),
(Instant.of_epoch_milli(2000), 'Bob', 53.1),
(Instant.of_epoch_milli(5000), 'Alice', 13.1),
(Instant.of_epoch_milli(3000), 'Bob', 3.1),
(Instant.of_epoch_milli(7000), 'Bob', 16.1),
(Instant.of_epoch_milli(10000), 'Alice', 20.1)
],
type_info=Types.ROW([Types.INSTANT(), Types.STRING(), Types.FLOAT()]))
table = t_env.from_data_stream(
ds,
Schema.new_builder()
.column_by_expression("ts", "CAST(f0 AS TIMESTAMP(3))")
.column("f1", DataTypes.STRING())
.column("f2", DataTypes.FLOAT())
.watermark("ts", "ts - INTERVAL '3' SECOND")
.build()
).alias("ts", "name", "price")
# define the sink
t_env.create_temporary_table(
'sink',
TableDescriptor.for_connector('print')
.schema(Schema.new_builder()
.column('name', DataTypes.STRING())
.column('total_price', DataTypes.FLOAT())
.build())
.build())
# define the over window operation
table = table.over_window(
Over.partition_by(col("name"))
.order_by(col("ts"))
.preceding(row_interval(2))
.following(CURRENT_ROW)
.alias('w')) \
.select(table.name, table.price.max.over(col('w')))
# submit for execution
table.execute_insert('sink') \
.wait()
