Hadoop interviews still come up for big-data platform, legacy-maintenance and migration roles: HDFS internals, high availability, YARN, MapReduce, file formats and the small-files problem, plus where Hadoop sits now that many teams run Spark on cloud object storage. As of November 17, 2026, Apache Hadoop's releases page lists 3.5.0 (April 2, 2026) and 3.4.3 (February 24, 2026) as current lines, and hdfs-default.xml still sets a 128 MB block size and three replicas. Below are 30 questions by level, five scenarios and how interviews test Hadoop.
HDFS facts from the Apache docs
| Setting or fact | Value (checked November 17, 2026) |
|---|---|
| dfs.blocksize | 134217728 bytes (128 MB) |
| dfs.replication | 3 |
| dfs.namenode.checkpoint.period | 3600 seconds between checkpoints |
| Erasure coding | RS(6,3) enabled by default; overhead of at most 50% versus 200% for 3x replication |
| Latest releases | 3.5.0 (April 2, 2026) and 3.4.3 (February 24, 2026) |
| Retired ecosystem projects | Sqoop moved to the Apache Attic in June 2021; Oozie is also in the Attic |
Fundamentals
- What are Hadoop's core modules? HDFS for distributed storage, YARN for resource management and scheduling, MapReduce for batch processing, and Hadoop Common for shared libraries.
- NameNode or DataNode? The NameNode holds the namespace and the map of blocks to DataNodes in memory; DataNodes store blocks on disk and send heartbeats and block reports to the NameNode.
- Is the Secondary NameNode a backup? No. It periodically merges the fsimage with the edit log into a new checkpoint, every 3,600 seconds by default. High availability needs a Standby NameNode.
- Why is the HDFS block size so large? Large blocks, 128 MB by default, keep seek time small relative to transfer time and keep the NameNode's metadata manageable. A file smaller than a block doesn't use a full block of disk.
- How do replication and rack awareness work? Each block is replicated three times by default. The default policy puts one replica on the writer's node, a second on a node in a different rack and a third on another node in that second rack, surviving the loss of a rack.
- Describe the HDFS write and read paths. On write, the client asks the NameNode for target DataNodes and streams the block through a pipeline of them, each acknowledging. On read, the client gets block locations and reads from the closest replica.
- How does NameNode high availability work? An active and one or more standby NameNodes share edits through JournalNodes, and ZooKeeper Failover Controllers trigger automatic failover. Hadoop 3 supports more than one standby.
- What is safe mode? A read-only state at startup until enough block replicas have been reported. Check it with
hdfs dfsadmin -safemode get; leaving it by force with missing blocks risks data loss. - How does YARN work? The ResourceManager schedules cluster resources, a NodeManager on each node runs containers, and an ApplicationMaster per application negotiates containers and tracks its tasks.
- Which YARN schedulers exist? FIFO, Capacity and Fair. The Capacity Scheduler, the default in Apache Hadoop, gives queues guaranteed shares with elastic use of idle capacity.
- Which HDFS commands do you use daily?
hdfs dfs -put,-lsand-du -hfor files,hdfs fsck / -files -blocks -locationsfor block health andhdfs dfsadmin -reportfor capacity and live nodes.
Intermediate
- Walk through a MapReduce job. The input format creates splits, map tasks emit key-value pairs, output is partitioned, sorted and spilled, the shuffle copies partitions to reducers, which merge and reduce them, and the output format writes results.
- InputSplit or block? A block is physical storage; an InputSplit is the logical chunk a mapper processes. Records that cross a block boundary are read whole by the split that starts them.
- What is a combiner, and when can't you use one? A map-side mini-reducer that cuts shuffle volume. It must be safe to run zero or more times, so it suits sums and maxima but not a direct average; combine sums and counts instead.
- What does a partitioner do? It decides which reducer receives each key; the default HashPartitioner hashes the key. Write a custom one to group keys deliberately or spread a hot key.
- How many mappers and reducers run? Mappers follow the number of input splits; reducers are set by the job, for example
mapreduce.job.reduces, and zero reducers makes a map-only job. - What is speculative execution? Launching a duplicate attempt of a slow task and keeping whichever finishes first. Disable it for tasks with external side effects.
- What is the distributed cache for? Shipping read-only files, such as lookup data or jars, to every node, which enables map-side joins with a small table.
- What is the small-files problem? Every file and block is an object in NameNode memory, and each small file can mean its own map task. Fix it with compaction jobs, container formats such as SequenceFile, ORC or Parquet, and combined input formats.
- Which file formats and codecs do you choose? Avro for row-oriented data with schema evolution; Parquet or ORC for columnar analytics with predicate pushdown. Gzip text isn't splittable; Snappy inside a container format usually is the practical choice.
- What is HDFS erasure coding? Hadoop 3's alternative to replication for warm and cold data. RS(6,3) stores six data and three parity blocks, using 9 blocks of disk where 3x replication of 6 blocks uses 18; an EC file's replication factor is always 1.
Experienced and senior
- How do you scale the NameNode? Size its heap to the number of files and blocks, fight small files, and use federation, with several namespaces, or router-based federation when one namespace isn't enough.
- How do you decommission a DataNode safely? Add it to the exclude file, run
hdfs dfsadmin -refreshNodes, and wait until it shows decommissioned, so its blocks are re-replicated first. - What does the balancer do? It moves blocks so disk usage evens out across DataNodes, for example after adding nodes:
hdfs balancer -threshold 10. - How is a Hadoop cluster secured? Kerberos for authentication, HDFS permissions and ACLs, Apache Ranger for central authorization policies, encryption zones for data at rest and wire encryption in transit.
- How do you size YARN containers? Set the memory each NodeManager offers (
yarn.nodemanager.resource.memory-mb) and per-task requests such asmapreduce.map.memory.mb, with the JVM heap below the container size to leave headroom. - Hadoop MapReduce or Spark? Spark keeps intermediate data in memory across a DAG of stages and runs on YARN against HDFS, so it is much faster for iterative and interactive work; MapReduce writes to disk between steps. See our Spark interview questions.
- What does the ecosystem look like in 2026? Hive is active, with 4.2.1 released in August 2026, while Sqoop and Oozie have been retired to the Apache Attic; teams replaced them with Spark JDBC reads, change-data-capture tools and orchestrators such as Airflow. See our Hive interview questions.
- How do you migrate Hadoop workloads to the cloud? Copy HDFS data to object storage with DistCp, move the Hive Metastore or adopt a table format such as Iceberg, run Spark on a managed service, and validate with counts and checksums before switching off the cluster.
- How do you debug a slow job? Check job counters for spills and records, look for one task running far longer than the rest (skew), check garbage collection and data locality, then fix the partitioning or the input files.
Five scenarios interviewers give
- The NameNode stays in safe mode after a restart. Blocks are missing because DataNodes are down or disks failed. Bring the nodes back and run fsck before forcing safe mode off.
- Under-replicated blocks spike after a node failure. Expected; HDFS re-replicates on its own. Watch progress and the network, and investigate if the count doesn't fall.
- One reducer runs for hours while the rest finish. Key skew. Use a custom partitioner, salt the hot key or pre-aggregate in a combiner.
- Some DataNodes are 95% full while others sit at 40%. Run the balancer and check the block placement policy and rack configuration.
- Streaming ingestion created millions of tiny files. Add a compaction job into ORC or Parquet and batch writes into larger files.
How interviews test Hadoop
Employers don't publish question banks, but big-data candidates commonly report concept questions like those above, a scenario on failure or performance, Spark and Hive questions alongside, and shell or Linux tasks; see our Linux interview questions. For broader pipeline design, see our data engineer interview questions.
Where CoPilot Interview fits
CoPilot Interview is an AI interview assistant whose free AI mock interview builds Hadoop and big data questions from your résumé and the job description, reads them aloud and scores each spoken answer out of 10. The browser demo at /try has 145 practice questions in 10 tracks, scored out of 10, with no signup. The free plan gives 15 minutes a day of live help in sessions of up to 15 minutes, with AI mock interviews included; Standard costs $24.99 a month and Pro $39.99. For data roles, see our guide to practicing data engineer interviews with AI. Live coding tasks, take-homes and online assessments usually forbid outside help, so prepare for those; in any live interview, follow your employer's rules.
Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly.
Frequently asked questions
Is Hadoop still relevant in 2026?
As a platform for new projects, less so: many teams run Spark on cloud object storage. But Apache Hadoop shipped 3.5.0 in April 2026, many enterprises still run clusters, and HDFS and YARN concepts underpin migration work.
What is the default HDFS block size?
128 MB (134217728 bytes), per hdfs-default.xml as of November 17, 2026. It is configurable per cluster and per file.
What is the difference between the Secondary NameNode and a Standby NameNode?
The Secondary NameNode only builds checkpoints; a Standby NameNode in an HA setup keeps its namespace current and takes over on failover.
What is the difference between Hadoop and Spark?
Hadoop is a storage and resource platform with MapReduce as its original engine; Spark is a faster processing engine that can run on Hadoop's YARN and read HDFS.
How do I prepare for a Hadoop interview?
Run a single-node cluster in Docker, load data, break a DataNode and watch re-replication, and write one MapReduce or Spark job with a custom partitioner.
Practice this interview before the real one
The AI mock interview builds questions from your resume and the job description, reads them aloud and scores each spoken answer out of 10. Included on the free plan.
Start a mock interview