Why it matters
Every large HDFS cluster is a multi-tenant environment. Finance runs nightly ETL, marketing runs analytics, data science experiments with a variety of workloads, and each of them tends to think their workload is the priority. Without quotas, the first team to overrun capacity forces every other team into contention. With quotas, capacity planning becomes a straightforward allocation exercise and cluster health becomes predictable.
Quotas also protect against operator mistakes. A misconfigured Spark job that writes a billion tiny files is a top cause of NameNode outages in the wild. A quota that caps the number of files under the destination directory catches this before it becomes a heap-pressure incident. That single guardrail justifies the operational overhead of managing quotas.
The architecture
Quotas are attached to directories in the NameNode inode tree. Each inode carries two optional quota fields: a namespace quota expressed as a count of files plus directories, and a space quota expressed as a byte budget. The NameNode tracks current usage against each quota at all times, updating counters incrementally as inodes are added and removed and as blocks are allocated or deleted.
The space quota multiplies user data by the replication factor. If a file has three replicas at one hundred megabytes each, it consumes three hundred megabytes of space quota. This means that if you lower the replication factor of an existing file, its space consumption drops proportionally. It also means that erasure-coded files consume space quota at their EC overhead ratio (roughly one and a half times user data) rather than the three times replication ratio.
How it works end to end
An operator sets a quota with hdfs dfsadmin -setQuota for namespace or -setSpaceQuota for storage. The NameNode records the quota on the target directory's inode and immediately begins tracking. Existing usage is computed on the fly by walking the subtree, which can take seconds for a large tree. The quota takes effect atomically once the walk completes.
Every mutation that could grow the quota consumption goes through a quota check. mkdir consumes one directory slot. create consumes one file slot plus the space that the initial block will need. allocateBlock consumes another block's worth of space quota. If any check exceeds the configured quota, the NameNode rejects the RPC with a DSQuotaExceededException or NSQuotaExceededException, and the client-side error propagates back to the user.
Quotas are inherited in the sense that a child directory's usage counts toward every ancestor's quota. If /team/finance/data has a quota and /team/finance has a larger quota, both are checked on any allocation under the deeper path. This lets you set a broad cap at the top level and finer caps for specific projects underneath.