Fuzzy dedup modifications (#687)

* Driver broadcast of large spark config variables Signed-off-by: Constantin M Adam <[email protected]> * Launch k8s spark cluster using spark native k8s API Signed-off-by: Constantin M Adam <[email protected]> * Add PyYAML to the base image dependencies Signed-off-by: Constantin M Adam <[email protected]> * Updated documentation for get_bcast_params() method Signed-off-by: Constantin M Adam <[email protected]> * Updated documentation Signed-off-by: Constantin M Adam <[email protected]> --------- Signed-off-by: Constantin M Adam <[email protected]>
IBM · Oct 11, 2024 · 2869eea · 2869eea
1 parent fa76928
commit 2869eea
Show file tree

Hide file tree

Showing 5 changed files with 95 additions and 16 deletions.
diff --git a/data-processing-lib/doc/spark-runtime.md b/data-processing-lib/doc/spark-runtime.md
@@ -41,9 +41,11 @@ of this parameter:
 
 ## Transforms
 
-* [SparkTransformRuntimeConfiguration](../spark/src/data_processing_spark/transform/runtime_configuration.py) allows
-    to configure transform to use PySpark
-
+* [SparkTransformRuntimeConfiguration](../spark/src/data_processing_spark/runtime/spark/runtime_configuration.py)
+  allows to configure transform to use PySpark. In addition to its base class
+  [TransformRuntimeConfiguration](../python//src/data_processing/runtime/runtime_configuration.py) features,
+  this class includes `get_bcast_params()` method to get very large configuration settings. Before starting the
+  transform execution, the Spark runtime will broadcast these settings to all the workers.
 
 ## Runtime
 

diff --git a/data-processing-lib/spark/pyproject.toml b/data-processing-lib/spark/pyproject.toml
@@ -13,7 +13,8 @@ authors = [
 dependencies = [
     "data-prep-toolkit==0.2.2.dev0",	
     "pyspark>=3.5.2",
-    "psutil>=6.0.0"
+    "psutil>=6.0.0",
+    "PyYAML>=6.0.2"
 ]
 
 [project_urls]

diff --git a/data-processing-lib/spark/src/data_processing_spark/runtime/spark/runtime_configuration.py b/data-processing-lib/spark/src/data_processing_spark/runtime/spark/runtime_configuration.py
@@ -10,6 +10,9 @@
 # limitations under the License.
 ################################################################################
 
+from typing import Any
+
+from data_processing.data_access import DataAccessFactoryBase
 from data_processing.runtime import TransformRuntimeConfiguration
 from data_processing.transform import TransformConfiguration
 from data_processing_spark.runtime.spark import DefaultSparkTransformRuntime
@@ -29,6 +32,16 @@ def __init__(
         super().__init__(transform_config=transform_config)
         self.runtime_class = runtime_class
 
+    def get_bcast_params(self, data_access_factory: DataAccessFactoryBase) -> dict[str, Any]:
+        """Allows retrieving and broadcasting to all the workers very large
+        configuration parameters, like the list of document IDs to remove for
+        fuzzy dedup, or the list of blocked web domains for block listing.  This
+        function is called by the spark runtime after spark initialization, and
+        before spark_context.parallelize()
+        :param data_access_factory - creates data_access object to download the large config parameter
+        """
+        return {}
+
     def create_transform_runtime(self) -> DefaultSparkTransformRuntime:
         """
         Create transform runtime with the parameters captured during apply_input_params()

diff --git a/data-processing-lib/spark/src/data_processing_spark/runtime/spark/transform_orchestrator.py b/data-processing-lib/spark/src/data_processing_spark/runtime/spark/transform_orchestrator.py
@@ -10,24 +10,69 @@
 # limitations under the License.
 ################################################################################
 
+import os
+import socket
 import time
 import traceback
 from datetime import datetime
 
+import yaml
 from data_processing.data_access import DataAccessFactoryBase
 from data_processing.transform import TransformStatistics
 from data_processing.utils import GB, get_logger
 from data_processing_spark.runtime.spark import (
+    SparkTransformExecutionConfiguration,
     SparkTransformFileProcessor,
     SparkTransformRuntimeConfiguration,
-    SparkTransformExecutionConfiguration,
 )
 from pyspark import SparkConf, SparkContext
+from pyspark.sql import SparkSession
 
 
 logger = get_logger(__name__)
 
 
+def _init_spark(runtime_config: SparkTransformRuntimeConfiguration) -> SparkSession:
+    server_port_https = int(os.getenv("KUBERNETES_SERVICE_PORT_HTTPS", "-1"))
+    if server_port_https == -1:
+        # running locally
+        spark_config = {"spark.driver.host": "127.0.0.1"}
+        return SparkSession.builder.appName(runtime_config.get_name()).config(map=spark_config).getOrCreate()
+    else:
+        # running in Kubernetes, use spark_profile.yml and
+        # environment variables for configuration
+        server_port = os.environ["KUBERNETES_SERVICE_PORT"]
+        master_url = f"k8s://https://kubernetes.default:{server_port}"
+
+        # Read Spark configuration profile
+        config_filepath = os.path.abspath(
+            os.path.join(os.getenv("SPARK_HOME"), "work-dir", "config", "spark_profile.yml")
+        )
+        with open(config_filepath, "r") as config_fp:
+            spark_config = yaml.safe_load(os.path.expandvars(config_fp.read()))
+        spark_config["spark.submit.deployMode"] = "client"
+
+        # configure the executor pods from template
+        executor_pod_template_file = os.path.join(
+            os.getenv("SPARK_HOME"),
+            "work-dir",
+            "src",
+            "templates",
+            "spark-executor-pod-template.yml",
+        )
+        spark_config["spark.kubernetes.executor.podTemplateFile"] = executor_pod_template_file
+        spark_config["spark.kubernetes.container.image.pullPolicy"] = "Always"
+
+        # Pass the driver IP address to the workers for callback
+        myservice_url = socket.gethostbyname(socket.gethostname())
+        spark_config["spark.driver.host"] = myservice_url
+        spark_config["spark.driver.bindAddress"] = "0.0.0.0"
+        spark_config["spark.decommission.enabled"] = True
+        logger.info(f"Launching Spark Session with configuration\n" f"{yaml.dump(spark_config, indent=2)}")
+        app_name = spark_config.get("spark.app.name", "my-spark-app")
+        return SparkSession.builder.master(master_url).appName(app_name).config(map=spark_config).getOrCreate()
+
+
 def orchestrate(
     runtime_config: SparkTransformRuntimeConfiguration,
     execution_configuration: SparkTransformExecutionConfiguration,
@@ -45,14 +90,17 @@ def orchestrate(
     logger.info(f"orchestrator started at {start_ts}")
     # create data access
     data_access = data_access_factory.create_data_access()
+    bcast_params = runtime_config.get_bcast_params(data_access_factory)
     if data_access is None:
         logger.error("No DataAccess instance provided - exiting")
         return 1
     # initialize Spark
-    conf = SparkConf().setAppName(runtime_config.get_name()).set("spark.driver.host", "127.0.0.1")
-    sc = SparkContext(conf=conf)
+    spark_session = _init_spark(runtime_config)
+    sc = spark_session.sparkContext
+    # broadcast
     spark_runtime_config = sc.broadcast(runtime_config)
     daf = sc.broadcast(data_access_factory)
+    spark_bcast_params = sc.broadcast(bcast_params)
 
     def process_partition(iterator):
         """
@@ -63,6 +111,7 @@ def process_partition(iterator):
         # local statistics dictionary
         statistics = TransformStatistics()
         # create transformer runtime
+        bcast_params = spark_bcast_params.value
         d_access_factory = daf.value
         runtime_conf = spark_runtime_config.value
         runtime = runtime_conf.create_transform_runtime()
@@ -77,8 +126,11 @@ def process_partition(iterator):
                 logger.debug(f"partition {f}")
                 # add additional parameters
                 transform_params = (
-                    runtime.get_transform_config(partition=int(f[1]), data_access_factory=d_access_factory,
-                                                 statistics=statistics))
+                    runtime.get_transform_config(
+                        partition=int(f[1]), data_access_factory=d_access_factory, statistics=statistics
+                    )
+                    | bcast_params
+                )
                 # create transform with partition number
                 file_processor.create_transform(transform_params)
                 first = False
@@ -128,7 +180,7 @@ def process_partition(iterator):
         memory = 0.0
         for i in range(executors.size()):
             memory += executors.toList().apply(i)._2()._1()
-        resources = {"cpus": cpus, "gpus": 0, "memory": round(memory/GB, 2), "object_store": 0}
+        resources = {"cpus": cpus, "gpus": 0, "memory": round(memory / GB, 2), "object_store": 0}
         input_params = runtime_config.get_transform_metadata() | execution_configuration.get_input_params()
         metadata = {
             "pipeline": execution_configuration.pipeline_id,
@@ -143,7 +195,8 @@ def process_partition(iterator):
             "execution_stats": {
                 "num partitions": num_partitions,
                 "execution time, min": round((time.time() - start_time) / 60, 3),
-            } | resources,
+            }
+            | resources,
             "job_output_stats": stats,
         }
         logger.debug(f"Saving job metadata: {metadata}.")

diff --git a/data-processing-lib/spark/src/data_processing_spark/runtime/spark/transform_runtime.py b/data-processing-lib/spark/src/data_processing_spark/runtime/spark/transform_runtime.py
@@ -34,19 +34,29 @@ def get_transform_config(
         """
         Get the dictionary of configuration that will be provided to the transform's initializer.
         This is the opportunity for this runtime to create a new set of configuration based on the
-        config/params provided to this instance's initializer.  This may include the addition
-        of new configuration data such as ray shared memory, new actors, etc, that might be needed and
-        expected by the transform in its initializer and/or transform() methods.
+        config/params provided to this instance's initializer.
+        :param partition - the partition assigned to this worker, needed by transforms like doc_id
         :param data_access_factory - data access factory class being used by the RayOrchestrator.
         :param statistics - reference to statistics actor
         :return: dictionary of transform init params
         """
         return self.params
 
+    def get_bcast_params(self, data_access_factory: DataAccessFactoryBase) -> dict[str, Any]:
+        """Allows retrieving and broadcasting to all the workers very large
+        configuration parameters, like the list of document IDs to remove for
+        fuzzy dedup, or the list of blocked web domains for block listing. This
+        function is called by the spark runtime after spark initialization, and
+        before spark_context.parallelize()
+        :param data_access_factory - creates data_access object to download the large config parameter
+        """
+        return {}
+
     def compute_execution_stats(self, stats: TransformStatistics) -> None:
         """
         Update/augment the given statistics object with runtime-specific additions/modifications.
+        This method does not return a value; the job execution statistics are generally reported
+        as metadata by the Spark Orchestrator.
         :param stats: output of statistics as aggregated across all calls to all transforms.
-        :return: job execution statistics.  These are generally reported as metadata by the Ray Orchestrator.
         """
-        pass
+        pass