Coordination Services and Network-Attached Storage

Coordination Services

Advisory lock: A lock that the system does not enforce; well-behaved processes check for lock ownership before proceeding, but nothing prevents code from ignoring it.
Cache invalidation: A message sent by a server to clients that hold cached copies of a resource, telling them their copy is no longer valid.
Chubby cell: A deployment unit of Google’s Chubby service, consisting of five replicas with one elected master; clients within a data center contact their local cell.
Coarse-grained lock: A lock held for a long period (minutes to hours) over a large resource such as a master role or an entire table, as opposed to a fine-grained lock held for milliseconds over a small resource.
Coordination service: A small, strongly consistent, highly available replicated store used by distributed systems for control-plane decisions such as leader election, locking, and configuration management.
Ephemeral node (znode): A ZooKeeper node that is automatically deleted when the client session that created it ends, used as the primary mechanism for failure detection.
etcd: A strongly consistent distributed key-value store using Raft consensus, serving as the authoritative state store for Kubernetes.
Fencing token: A monotonically increasing number issued with each lock grant; the protected resource rejects requests carrying a token lower than the highest it has seen, preventing a stale lock holder from corrupting shared state.
Fine-grained lock: A lock held briefly (milliseconds) to protect a small resource such as a single record; not suitable for consensus-backed coordination services due to high acquisition frequency.
Grace period: In Chubby, the window of time after a master failover during which clients may reconnect and re-establish their sessions before the new master releases their locks.
Lease: A time-bounded grant from a server to a client; the client may rely on the grant (cached data, lock ownership) until the lease expires, and must renew it periodically to maintain it.
Leader election: The coordination pattern by which one replica among several atomically claims a well-known name in a coordination service and holds it as long as its session is alive.
Linearizable read: A read that reflects the most recently committed write, as if the entire system had a single consistent view at that instant; the strongest read consistency guarantee.
Persistent node (znode): A ZooKeeper node that survives client disconnection and remains until explicitly deleted.
Sequential node (znode): A ZooKeeper node created with a monotonically increasing integer appended to its name, used to implement ordered lock queues and prevent the thundering herd problem.
Service discovery: The coordination pattern by which running service instances register their addresses under ephemeral keys so that clients can find currently available instances.
Session timeout: In ZooKeeper, the negotiated period within which a client must send heartbeats; if the timeout elapses without a heartbeat, the session ends and all ephemeral nodes created by that session are deleted.
Thundering herd problem: A condition where a single state change simultaneously wakes many waiting clients, causing a burst of requests that overwhelms the system even though only one client can make progress.
Watch: In ZooKeeper and etcd, a one-shot notification that fires when a specified znode or key changes; the client must re-register the watch after each firing.
ZooKeeper: An open-source coordination service from Yahoo that provides a hierarchical namespace of znodes and primitives for building locks, leader election, and other coordination patterns.

Network-Attached Storage: General Concepts

Access transparency: The property that remote files are accessible via the same system calls as local files, with no changes required in applications.
Close-to-open consistency: A caching model in which a client checks the server for updates when a file is opened; stale reads are possible between opens; used by NFS.
Mount point: A location in a directory tree where a different file system (local or remote) is attached, making its contents accessible under that path.
Read-ahead: A client-side optimization that speculatively fetches blocks beyond what the application has requested, anticipating sequential access and hiding network latency.
Referral: A response from a running server directing a client to contact a different server for a resource; used for planned migrations and namespace federation, not failure recovery.
Session semantics: A caching model in which a client’s changes to a file are not visible to other clients until the file is closed and uploaded; last writer wins on conflict; used by AFS and Coda.
Stateless server: A server that holds no information about client activity between requests; each request carries all necessary context, making crash recovery trivial.
Virtual File System (VFS): A kernel abstraction layer that defines a standard interface for file system operations; allows local and remote file systems to coexist and be mounted into a single directory tree.
Write-behind: A caching policy in which modifications are accumulated locally and sent to the server in batches; more efficient than write-through but leaves the server with stale data in the interim.
Write-through: A caching policy in which every write is sent immediately to the server, keeping the server’s copy always current.

NFS

File handle (NFS): An opaque server-generated identifier for a file or directory, used by the client in all subsequent requests; persists across server restarts to support stateless operation.
Network Lock Manager (NLM): A separate service added alongside NFS to provide advisory file locking, a stateful function that could not be part of the stateless NFS protocol itself.
NFS (Network File System): A stateless, RPC-based distributed file system designed for simplicity and cross-platform interoperability; uses file handles and timestamp-based cache validation.
Silly rename: An NFS client workaround for the lack of open-file reference counting: a file to be deleted while locally open is renamed to a hidden name and deleted only when the last local file descriptor closes.
Timestamp validation: The NFS mechanism for cache consistency; the client compares the server’s file modification time against its cached copy and discards the cache if the server’s version is newer.

AFS

AFS callback promise: A server-to-client guarantee that the server will send a revocation if a cached file is modified, allowing the client to use its local copy indefinitely without polling.
AFS callback revocation: A notification sent by an AFS server to all clients that have cached a file, telling them to invalidate their copy because the file has been modified.
Andrew File System (AFS): A distributed file system using whole-file caching and server callbacks to achieve high scalability; enforces a uniform global namespace under /afs.
Uniform global namespace: An AFS property ensuring that the path to any file is identical on every client machine, in contrast to NFS where administrators mount remote directories at arbitrary local paths.
Volume (AFS/Coda): An administrative unit of the AFS and Coda file systems, typically corresponding to a user’s home directory or a software repository; can be moved between servers transparently via referrals.
Whole-file caching: The AFS model in which an entire file is downloaded to the client’s local disk on open; reads and writes operate on the local copy until the file is closed.

Coda

Accessible Volume Storage Group (AVSG): The subset of a VSG’s servers that a Coda client can currently reach; reads are served from any AVSG member and writes go to all of them simultaneously.
Client modification log (CML): In Coda, a log of file system operations (store, create, remove, rename) recorded during disconnected operation; actual file data remains in the local disk cache and is uploaded when the CML is replayed on reconnection.
Coda: A distributed file system extending AFS to support disconnected operation via a client modification log, conflict detection on reconnection, and cache hoarding.
Disconnected operation: A Coda operating mode in which the client works entirely from its local disk cache when no server is reachable, recording operations in the CML for later replay.
Hoarding: In Coda, the pre-population of the local cache with user-specified files before going offline to ensure they are available during disconnected operation.
Volume Storage Group (VSG): The set of servers that host a replicated Coda volume; the client checks that all accessible VSG members have consistent versions of a file before opening it.

SMB

Compound RPC: An NFSv4 and SMB 2 feature that packs multiple operations into a single network message, reducing round-trip latency for common operation sequences.
Delegation (NFSv4): A server-granted right allowing a client to cache reads or writes for a file without contacting the server; the server recalls the delegation if another client creates a conflict.
Distributed File System (DFS): A Microsoft namespace service, separate from SMB itself, that maps logical paths to physical server locations and issues referrals so clients are transparently redirected when content moves; predates SMB 2 and works across SMB versions.
Durable handle: An SMB 2 feature that preserves open file state across brief client disconnections, allowing the client to resume without reopening files or re-establishing locks.
Oplock break: A server-to-client message in SMB that revokes a previously granted opportunistic lock, requiring the client to flush any cached writes to the server before the conflicting access can proceed.
Opportunistic lock (oplock): An SMB server-granted caching right; the server monitors access and sends an oplock break to the caching client when a conflict arises, requiring it to flush cached writes before the competing open proceeds.
SMB (Server Message Block): Microsoft’s stateful, connection-oriented file sharing protocol; tracks open files, locks, and sessions at the server and enforces Windows file-sharing semantics.
SMB Multichannel: An SMB 3 feature that allows a single session to use multiple network interfaces simultaneously, increasing throughput and providing path redundancy.
Transparent Failover: An SMB 3 feature that allows a client connected to a clustered file server to survive the failure of one cluster node without losing open files or locks, because session state is shared across cluster nodes.