We use analytics and advertising cookies to understand how the site is used and whether our ads on Facebook and Instagram work. They are set only if you accept. See our Privacy Policy for details.
AWS Glue comes up in data engineering and analytics interviews as the serverless ETL and catalog layer of an S3 data lake. Expect scenarios: a crawler that splits one dataset into many tables, a job that reprocesses yesterday's files, a Spark job that runs out of memory, a query that scans every partition.
Each question below is a scenario with four plausible answers and an explanation.
Answer a real AWS interview question before you browse the Glue list. No signup needed.
A content management application runs on EC2 instances in an Auto Scaling group across three Availability Zones. Every instance must read and write the same set of user-uploaded files through a standard file system interface (POSIX), and files must remain available if an AZ fails. Which storage solution should a solutions architect choose?
Practice in a session
10 questions from AWS, one after another, then your score and streak.
The Data Catalog and crawlers, ETL jobs (Spark, Python shell, Ray) and worker types, DynamicFrames vs Spark DataFrames, job bookmarks for incremental loads, partitioning and pushdown predicates, triggers and workflows, Glue Data Quality, and how Glue fits with Athena, Lake Formation and EMR.
State that Glue keeps between runs of a job so the next run processes only data it has not seen - new files in S3, or new rows from a JDBC source with bookmark keys. The script must initialise and commit the job for the bookmark to advance, and you can reset or pause it to reprocess.
A DynamicFrame is Glue's self-describing collection of records that tolerates inconsistent schemas, recording a field that has several types as a choice you resolve with resolveChoice. A Spark DataFrame needs one schema up front. Glue converts between them with toDF and DynamicFrame.fromDF, so you can use Spark's API where you need it.
Glue is serverless Spark for ETL, billed per DPU-hour with no cluster to manage. EMR gives you control of the cluster, more frameworks (Spark, Hive, Presto/Trino, Flink) and cost levers such as instance choice and Spot for large or long-running jobs. Interviewers look for the operations-vs-control trade-off.