Does ConcurrentDictionary.GetOrAdd run the factory only once?

No. The factory runs outside the internal lock so your delegate cannot deadlock the map. Two threads can both miss, both run the factory, and only one value is stored. The loser is discarded. Do not put side effects in the factory.

When do I still need lock with ConcurrentDictionary?

When you mutate a value in place (List.Add on a shared Report), coordinate two keys, or must run an expensive factory exactly once. Dictionary methods are thread-safe; your TValue and your factory are not.

Is ConcurrentDictionary a cache?

It is a thread-safe map. It has no TTL, no size bound, and no stampede control unless you add them. For tenant fee schedules I usually want IMemoryCache plus a versioned key.

When should I use FrozenDictionary instead of ConcurrentDictionary?

When the map is built at startup and only swapped as a whole. ToFrozenDictionary is a faster reader, not a concurrent writer.

September 7, 20264 min readBy Muhammad Shahid

C# ConcurrentDictionary vs lock: GetOrAdd Can Run Twice

ConcurrentDictionary is a thread-safe map, not a cache policy. GetOrAdd’s factory can run twice, keys must include tenant id, and a lock is still right when you mutate a value in place.

Part of Async & Threading

Quick answers

Does ConcurrentDictionary.GetOrAdd run the factory only once?: No. The factory runs outside the internal lock so your delegate cannot deadlock the map. Two threads can both miss, both run the factory, and only one value is stored. The loser is discarded. Do not put side effects in the factory.
When do I still need lock with ConcurrentDictionary?: When you mutate a value in place (List.Add on a shared Report), coordinate two keys, or must run an expensive factory exactly once. Dictionary methods are thread-safe; your TValue and your factory are not.
Is ConcurrentDictionary a cache?: It is a thread-safe map. It has no TTL, no size bound, and no stampede control unless you add them. For tenant fee schedules I usually want IMemoryCache plus a versioned key.
When should I use FrozenDictionary instead of ConcurrentDictionary?: When the map is built at startup and only swapped as a whole. ToFrozenDictionary is a faster reader, not a concurrent writer.

A Dictionary<TKey,TValue> is not safe for two threads at once. ConcurrentDictionary lets many threads read and write the map without your own lock. Thread-safe means the map operations will not corrupt the buckets. It does not run your factory once, expire old values, or put tenant id in the key.

Multi-tenancy means many customers share one app; each must only see their data. Thread-safe is not tenant-safe.

New to this → stay here. Merging a PR → wrong vs right. On-call / interview → GetOrAdd twice · tenant keys · if an interviewer asks.

Terms used here: Factory = the lambda GetOrAdd calls to create a missing value. Stampede = many threads miss the same key at once. Single-flight = only one expensive load runs; others wait for that result. IMemoryCache = cache with size and TTL; use it when you need expiration.

Smallest example

var map = new ConcurrentDictionary<string, int>();
map.TryAdd("a", 1);
map.TryGetValue("a", out var n);

You do not wrap TryGetValue in your own lock “to be safe.”

GetOrAdd: the factory is not atomic

Microsoft’s docs: the valueFactory runs outside the dictionary lock so unknown code cannot block every writer.

Thread A: miss → OpenConnection(id) ──────────────────► discard (leaked socket)
Thread B: miss → OpenConnection(id) → wins insert

var map = new ConcurrentDictionary<string, SqlConnection>();

var connection = map.GetOrAdd(tenantId, id =>
{
    // Can run twice for the same id under a stampede
    return OpenConnection(id);
});

Do not put side effects, logging “initialized,” or HttpClient.Send in that factory.

Single-flight with Lazy

var map = new ConcurrentDictionary<string, Lazy<FeeSchedule>>();

FeeSchedule GetSchedule(string tenantId) =>
    map.GetOrAdd(
            tenantId,
            id => new Lazy<FeeSchedule>(() => LoadFromSql(id)))
        .Value;

You may still allocate extra Lazy<T> objects. Only one LoadFromSql runs per key with default LazyThreadSafetyMode.ExecutionAndPublication. .Value blocks other threads until the winner finishes.

Advanced: Lazy<Task<T>> — a faulted task stays cached until you remove the key. Do not capture a request CancellationToken into the factory (first caller cancels → everyone shares the canceled task).

Wrong vs right

I would reject a key of "fees" in a multi-tenant app and List.Add on a value stored in the map.

// Wrong — last clinic wins
map.GetOrAdd("fees", _ => Load(clinicId));

// Right
map.GetOrAdd($"{clinicId}:{version}", _ => Load(clinicId, version));

var report = map.GetOrAdd(key, _ => new Report());
report.Lines.Add(line); // List<T> is not thread-safe

AddOrUpdate vs GetOrAdd

Comparison: GetOrAdd versus AddOrUpdate versus TryAdd.

Method	Use when
`GetOrAdd`	Read-through: return existing or insert
`AddOrUpdate`	Insert or replace using an update delegate
`TryAdd` / `TryUpdate` / `TryRemove`	You need to branch on success

AddOrUpdate’s delegates can also run more than once. Treat them as pure. For a hot counter, Interlocked is cheaper.

Tenant keys and unbounded growth

A key of "fees" is thread-safe and still a cross-tenant leak. Include tenant id (and a version) in the key. Thread-safe is not tenant-safe.

A singleton dictionary grows until the process dies. IMemoryCache with size limits is the product cache. ConcurrentDictionary is for coordination: single-flight, idempotency keys, “is this message in flight.”

FrozenDictionary for read-mostly maps

A fee schedule loaded at startup and never mutated is not a ConcurrentDictionary problem. using System.Collections.Frozen; then ToFrozenDictionary — faster reader, not a concurrent writer. If ops publish a new schedule, build a new frozen map and swap the field.

Common mistakes

Factory that hits SQL with a request-scoped DbContext captured into a singleton map
Caching HttpContext or AsyncLocal user into a static dictionary
Using Count in a hot loop

What this is not

Broader primitives: lock vs Monitor. Staff-loop cache questions: expert C# interviews. Topic map: async & threading hub.

If an interviewer asks

Thread-safe collections; why Singleton services must be thread-safe; GetOrAdd running twice.

Strong answer: “Thread-safe structure, not a cache policy. Factory can run twice. Keys include tenant.”

. Bring the key shape and the DI lifetime of the dictionary.