Run Spark jobs using Amazon EMR Serverless

Download the first 100 station files for 2024.

$ mkdir -p ~/gsod && cd ~/gsod
aws s3 ls s3://noaa-gsod-pds/2024/ --no-sign-request | head -100 | awk '{print $NF}' > files.txt
while read f; do
    aws s3 cp "s3://noaa-gsod-pds/2024/$f" "$f" --no-sign-request --only-show-errors
done < files.txt
ls | wc -l

The command downloads approximately 100 CSV files totaling about 7–8 MB.

Upload the files to the access point under the gsod/2024/ prefix. Replace access-point-alias with your access point alias.

$ aws s3 cp ~/gsod/ "s3://access-point-alias/gsod/2024/" --recursive --exclude "files.txt" --only-show-errors

Save the following script to a file named gsod_monthly.py.

# gsod_monthly.py
import sys
from pyspark.sql import SparkSession
from pyspark.sql import functions as F

INPUT_PATH, OUTPUT_PATH = sys.argv[1], sys.argv[2]

# GSOD sentinels for missing data
TEMP_SENTINEL = 9999.9
PRCP_SENTINEL = 99.99

spark = SparkSession.builder.appName("gsod-monthly-summary").getOrCreate()

raw = spark.read.option("header", True).csv(INPUT_PATH)

cleaned = (raw
    .select(
        F.col("STATION").alias("station"),
        F.col("NAME").alias("station_name"),
        F.col("LATITUDE").cast("double").alias("lat"),
        F.col("LONGITUDE").cast("double").alias("lon"),
        F.to_date("DATE", "yyyy-MM-dd").alias("date"),
        F.col("TEMP").cast("double").alias("temp_f"),
        F.col("PRCP").cast("double").alias("prcp_in"),
        F.col("FRSHTT").alias("frshtt"),
    )
    .filter(F.col("temp_f") != TEMP_SENTINEL)
    .withColumn("month", F.date_format("date", "yyyy-MM"))
    .withColumn(
        "prcp_in",
        F.when(F.col("prcp_in") == PRCP_SENTINEL, None).otherwise(F.col("prcp_in")),
    )
    # FRSHTT is a 6-char bitfield: Fog, Rain, Snow, Hail, Thunder, Tornado.
    # Check only positions 2-4 (Rain, Snow, Hail) for precipitation events.
    .withColumn(
        "had_precip_event",
        F.when(F.col("frshtt").substr(2, 3).rlike("1"), 1).otherwise(0),
    )
)

monthly = (cleaned
    .groupBy("station", "station_name", "lat", "lon", "month")
    .agg(
        F.avg("temp_f").alias("avg_temp_f"),
        F.min("temp_f").alias("min_temp_f"),
        F.max("temp_f").alias("max_temp_f"),
        F.sum("prcp_in").alias("total_prcp_in"),
        F.sum("had_precip_event").alias("precip_event_days"),
        F.count("*").alias("observation_days"),
    )
)

(monthly.write
    .mode("overwrite")
    .partitionBy("month")
    .parquet(OUTPUT_PATH))

spark.stop()

Upload the script to the access point under the scripts/ prefix.

$ aws s3 cp gsod_monthly.py "s3://access-point-alias/scripts/gsod_monthly.py"

Create a Spark application using a recent Amazon EMR release.

$ aws emr-serverless create-application \
    --name fsxn-emr-app --type SPARK --release-label emr-7.0.0

Note the applicationId in the response.

Start the application. Starting pre-warms a small pool of workers so the first job runs without a cold-start delay.

$ aws emr-serverless start-application --application-id application-id

Wait for the state to become STARTED.

$ aws emr-serverless get-application --application-id application-id \
    --query 'application.state'

Save the job driver configuration as job-driver.json. Replace the placeholders.

{
    "sparkSubmit": {
        "entryPoint": "s3://access-point-alias/scripts/gsod_monthly.py",
        "entryPointArguments": [
            "s3://access-point-alias/gsod/2024/",
            "s3://access-point-alias/gsod-monthly/"
        ],
        "sparkSubmitParameters": "--conf spark.executor.cores=2 --conf spark.executor.memory=4g --conf spark.driver.cores=2 --conf spark.driver.memory=4g --conf spark.executor.instances=2"
    }
}

Save the following monitoring configuration as job-config.json. It sends driver and executor logs to CloudWatch Logs.

{
    "monitoringConfiguration": {
        "cloudWatchLoggingConfiguration": {
            "enabled": true,
            "logGroupName": "/aws/emr-serverless/fsxn-emr-app"
        }
    }
}

Submit the job.

$ aws emr-serverless start-job-run \
    --application-id application-id \
    --execution-role-arn arn:aws:iam::account-id:role/fsxn-emr-job-role \
    --name gsod-monthly \
    --job-driver file://job-driver.json \
    --configuration-overrides file://job-config.json

Note the jobRunId in the response.

Poll the job status. The job transitions from SCHEDULED to RUNNING to SUCCESS.

$ aws emr-serverless get-job-run \
    --application-id application-id \
    --job-run-id job-run-id \
    --query 'jobRun.state'

List the output partitions.

$ aws s3 ls "s3://access-point-alias/gsod-monthly/" --recursive

You should see one Parquet file per month=YYYY-MM/ partition plus a _SUCCESS marker at the root.

Read a partition locally to verify the content.

$ aws s3 cp "s3://access-point-alias/gsod-monthly/month=2024-06/" . \
    --recursive --exclude "_SUCCESS"
python3 -c "import pyarrow.parquet as pq; \
    t = pq.read_table(next(__import__('glob').iglob('*.parquet'))); \
    print(t.schema); print(t.to_pandas().head())"

The output schema includes station, station_name, lat, lon, avg_temp_f, min_temp_f, max_temp_f, total_prcp_in, precip_event_days, and observation_days.