Read Isolation, Consistency, and Recency
Isolation Guarantees
Read Uncommitted
Depending on the read concern, clients can see the results of writes before the writes are durable:
Regardless of a write's write concern, other clients using
"local"
or"available"
read concern can see the result of a write operation before the write operation is acknowledged to the issuing client.Clients using
"local"
or"available"
read concern can read data which may be subsequently rolled back during replica set failovers.
For operations in a multi-document transaction, when a transaction commits, all data changes made in the transaction are saved and visible outside the transaction. That is, a transaction will not commit some of its changes while rolling back others.
Until a transaction commits, the data changes made in the transaction are not visible outside the transaction.
However, when a transaction writes to multiple shards, not all
outside read operations need to wait for the result of the committed
transaction to be visible across the shards. For example, if a
transaction is committed and write 1 is visible on shard A but write
2 is not yet visible on shard B, an outside read at read concern
"local"
can read the results of write 1 without
seeing write 2.
Read uncommitted is the default isolation level and applies to
mongod
standalone instances as well as to replica sets and
sharded clusters.
Read Uncommitted And Single Document Atomicity
Write operations are atomic with respect to a single document; i.e. if a write is updating multiple fields in the document, a read operation will never see the document with only some of the fields updated. However, although a client may not see a partially updated document, read uncommitted means that concurrent read operations may still see the updated document before the changes are made durable.
With a standalone mongod
instance, a set of read and write
operations to a single document is serializable. With a replica set,
a set of read and write operations to a single document is serializable
only in the absence of a rollback.
Read Uncommitted And Multiple Document Write
When a single write operation (e.g.
db.collection.updateMany()
) modifies multiple documents,
the modification of each document is atomic, but the operation as a
whole is not atomic.
When performing multi-document write operations, whether through a single write operation or multiple write operations, other operations may interleave.
For situations that require atomicity of reads and writes to multiple documents (in a single or multiple collections), MongoDB supports distributed transactions, including transactions on replica sets and sharded clusters.
For more information, see transactions
Important
In most cases, a distributed transaction incurs a greater performance cost over single document writes, and the availability of distributed transactions should not be a replacement for effective schema design. For many scenarios, the denormalized data model (embedded documents and arrays) will continue to be optimal for your data and use cases. That is, for many scenarios, modeling your data appropriately will minimize the need for distributed transactions.
For additional transactions usage considerations (such as runtime limit and oplog size limit), see also Production Considerations.
Without isolating the multi-document write operations, MongoDB exhibits the following behavior:
Non-point-in-time read operations. Suppose a read operation begins at time t 1 and starts reading documents. A write operation then commits an update to one of the documents at some later time t 2. The reader may see the updated version of the document, and therefore does not see a point-in-time snapshot of the data.
Non-serializable operations. Suppose a read operation reads a document d 1 at time t 1 and a write operation updates d 1 at some later time t 3. This introduces a read-write dependency such that, if the operations were to be serialized, the read operation must precede the write operation. But also suppose that the write operation updates document d 2 at time t 2 and the read operation subsequently reads d 2 at some later time t 4. This introduces a write-read dependency which would instead require the read operation to come after the write operation in a serializable schedule. There is a dependency cycle which makes serializability impossible.
Reads may miss matching documents that are updated during the course of the read operation.
Cursor Snapshot
MongoDB cursors can return the same document more than once in some situations. As a cursor returns documents, other operations may interleave with the query. If one of these operations changes the indexed field on the index used by the query, then the cursor could return the same document more than once.
Queries that use unique indexes can, in some cases, return duplicate values. If a cursor using a unique index interleaves with a delete and insert of documents sharing the same unique value, the cursor may return the same unique value twice from different documents.
Consider using read isolation. To learn more, see
Read Concern "snapshot"
.
Monotonic Writes
MongoDB provides monotonic write guarantees, by default, for
standalone mongod
instances and replica set.
For monotonic writes and sharded clusters, see Causal Consistency.
Real Time Order
For read and write operations on the primary, issuing read operations
with "linearizable"
read concern and write operations
with "majority"
write concern enables multiple threads
to perform reads and writes on a single document as if a single thread
performed these operations in real time; that is, the corresponding
schedule for these reads and writes is considered linearizable.
Causal Consistency
If an operation logically depends on a preceding operation, there is a causal relationship between the operations. For example, a write operation that deletes all documents based on a specified condition and a subsequent read operation that verifies the delete operation have a causal relationship.
With causally consistent sessions, MongoDB executes causal operations in an order that respect their causal relationships, and clients observe results that are consistent with the causal relationships.
Client Sessions and Causal Consistency Guarantees
To provide causal consistency, MongoDB enables causal consistency
in client sessions. A causally consistent session denotes that the
associated sequence of read operations with "majority"
read concern and write operations with "majority"
write
concern have a causal relationship that is reflected by their ordering.
Applications must ensure that only one thread at a time executes these
operations in a client session.
For causally related operations:
A client starts a client session.
Important
Client sessions only guarantee causal consistency for:
Read operations with
"majority"
; i.e. the return data has been acknowledged by a majority of the replica set members and is durable.Write operations with
"majority"
write concern; i.e. the write operations that request acknowledgment that the operation has been applied to a majority of the replica set's voting members.
For more information on causal consistency and various read and write concerns, see Causal Consistency and Read and Write Concerns.
As the client issues a sequence of read with
"majority"
read concern and write operations (with"majority"
write concern), the client includes the session information with each operation.For each read operation with
"majority"
read concern and write operation with"majority"
write concern associated with the session, MongoDB returns the operation time and the cluster time, even if the operation errors. The client session keeps track of the operation time and the cluster time.Note
MongoDB does not return the operation time and the cluster time for unacknowledged (
w: 0
) write operations. Unacknowledged writes do not imply any causal relationship.Although, MongoDB returns the operation time and the cluster time for read operations and acknowledged write operations in a client session, only the read operations with
"majority"
read concern and write operations with"majority"
write concern can guarantee causal consistency. For details, see Causal Consistency and Read and Write Concerns.The associated client session tracks these two time fields.
Note
Operations can be causally consistent across different sessions. MongoDB drivers and
mongosh
provide the methods to advance the operation time and the cluster time for a client session. So, a client can advance the cluster time and the operation time of one client session to be consistent with the operations of another client session.
Causal Consistency Guarantees
The following table lists the causal consistency guarantees provided by
causally consistent sessions for read operations with
"majority"
read concern and
write operations with "majority"
write concern.
Guarantees | Description |
---|---|
Read your writes | Read operations reflect the results of write
operations that precede them. |
Monotonic reads | Read operations do not return results that correspond to an earlier state of the data than a preceding read operation. For example, if in a session:
then read 2 cannot return results of write 1. |
Monotonic writes | Write operations that must precede other writes are executed before those other writes. For example, if write 1 must precede write 2 in a session, the state of the data at the time of write 2 must reflect the state of the data post write 1. Other writes can interleave between write 1 and write write 2, but write 2 cannot occur before write 1. |
Writes follow reads | Write operations that must occur after read operations are
executed after those read operations. That is, the state of the
data at the time of the write must incorporate the state of the
data of the preceding read operations. |
Read Preference
These guarantees hold across all members of the MongoDB deployment. For
example, if, in a causally consistent session, you issue a write with
"majority"
write concern followed by a read that reads
from a secondary (i.e. read preference secondary
) with
"majority"
read concern, the read operation will reflect
the state of the database after the write operation.
Isolation
Operations within a causally consistent session are not isolated from operations outside the session. If a concurrent write operation interleaves between the session's write and read operations, the session's read operation may return results that reflect a write operation that occurred after the session's write operation.
MongoDB Drivers
Tip
Applications must ensure that only one thread at a time executes these operations in a client session.
Clients require MongoDB drivers updated for MongoDB 3.6 or later:
Java 3.6+ Python 3.6+ C 1.9+ Go 1.8+ | C# 2.5+ Node 3.0+ Ruby 2.5+ Rust 2.1+ Swift 1.2+ | Perl 2.0+ PHPC 1.4+ Scala 2.2+ C++ 3.6.6+ |
Examples
Important
Causally consistent
sessions can only guarantee causal consistency for reads with
"majority"
read concern and writes with
"majority"
write concern.
Consider a collection items
that maintains the current and
historical data for various items. Only the historical data has a
non-null end
date. If the sku
value for an item changes, the
document with the old sku
value needs to be updated with the
end
date, after which the new document is inserted with the current
sku
value. The client can use a causally consistent session to
ensure that the update occurs before the insert.
➤ Use the Select your language drop-down menu in the upper-right to set the language of this example.
/* Use a causally-consistent session to run some operations. */ wc = mongoc_write_concern_new (); mongoc_write_concern_set_wmajority (wc, 1000); mongoc_collection_set_write_concern (coll, wc); rc = mongoc_read_concern_new (); mongoc_read_concern_set_level (rc, MONGOC_READ_CONCERN_LEVEL_MAJORITY); mongoc_collection_set_read_concern (coll, rc); session_opts = mongoc_session_opts_new (); mongoc_session_opts_set_causal_consistency (session_opts, true); session1 = mongoc_client_start_session (client, session_opts, &error); if (!session1) { fprintf (stderr, "couldn't start session: %s\n", error.message); goto cleanup; } /* Run an update_one with our causally-consistent session. */ update_opts = bson_new (); res = mongoc_client_session_append (session1, update_opts, &error); if (!res) { fprintf (stderr, "couldn't add session to opts: %s\n", error.message); goto cleanup; } query = BCON_NEW ("sku", "111"); update = BCON_NEW ("$set", "{", "end", BCON_DATE_TIME (bson_get_monotonic_time ()), "}"); res = mongoc_collection_update_one (coll, query, update, update_opts, NULL, /* reply */ &error); if (!res) { fprintf (stderr, "update failed: %s\n", error.message); goto cleanup; } /* Run an insert with our causally-consistent session */ insert_opts = bson_new (); res = mongoc_client_session_append (session1, insert_opts, &error); if (!res) { fprintf (stderr, "couldn't add session to opts: %s\n", error.message); goto cleanup; } insert = BCON_NEW ("sku", "nuts-111", "name", "Pecans", "start", BCON_DATE_TIME (bson_get_monotonic_time ())); res = mongoc_collection_insert_one (coll, insert, insert_opts, NULL, &error); if (!res) { fprintf (stderr, "insert failed: %s\n", error.message); goto cleanup; }
using (var session1 = client.StartSession(new ClientSessionOptions { CausalConsistency = true })) { var currentDate = DateTime.UtcNow.Date; var items = client.GetDatabase( "test", new MongoDatabaseSettings { ReadConcern = ReadConcern.Majority, WriteConcern = new WriteConcern( WriteConcern.WMode.Majority, TimeSpan.FromMilliseconds(1000)) }) .GetCollection<BsonDocument>("items"); items.UpdateOne(session1, Builders<BsonDocument>.Filter.And( Builders<BsonDocument>.Filter.Eq("sku", "111"), Builders<BsonDocument>.Filter.Eq("end", BsonNull.Value)), Builders<BsonDocument>.Update.Set("end", currentDate)); items.InsertOne(session1, new BsonDocument { {"sku", "nuts-111"}, {"name", "Pecans"}, {"start", currentDate} }); }
// Example 1: Use a causally consistent session to ensure that the update occurs before the insert. ClientSession session1 = client.startSession(ClientSessionOptions.builder().causallyConsistent(true).build()); Date currentDate = new Date(); MongoCollection<Document> items = client.getDatabase("test") .withReadConcern(ReadConcern.MAJORITY) .withWriteConcern(WriteConcern.MAJORITY.withWTimeout(1000, TimeUnit.MILLISECONDS)) .getCollection("test"); items.updateOne(session1, eq("sku", "111"), set("end", currentDate)); Document document = new Document("sku", "nuts-111") .append("name", "Pecans") .append("start", currentDate); items.insertOne(session1, document);
async with await client.start_session(causal_consistency=True) as s1: current_date = datetime.datetime.today() items = client.get_database( "test", read_concern=ReadConcern("majority"), write_concern=WriteConcern("majority", wtimeout=1000), ).items await items.update_one( {"sku": "111", "end": None}, {"$set": {"end": current_date}}, session=s1 ) await items.insert_one( {"sku": "nuts-111", "name": "Pecans", "start": current_date}, session=s1 )
my $s1 = $conn->start_session({ causalConsistency => 1 }); $items = $conn->get_database( "test", { read_concern => { level => 'majority' }, write_concern => { w => 'majority', wtimeout => 10000 }, } )->get_collection("items"); $items->update_one( { sku => 111, end => undef }, { '$set' => { end => $current_date} }, { session => $s1 } ); $items->insert_one( { sku => "nuts-111", name => "Pecans", start => $current_date }, { session => $s1 } );
$items = $client->selectDatabase( 'test', [ 'readConcern' => new \MongoDB\Driver\ReadConcern(\MongoDB\Driver\ReadConcern::MAJORITY), 'writeConcern' => new \MongoDB\Driver\WriteConcern(\MongoDB\Driver\WriteConcern::MAJORITY, 1000), ], )->items; $s1 = $client->startSession( ['causalConsistency' => true], ); $currentDate = new \MongoDB\BSON\UTCDateTime(); $items->updateOne( ['sku' => '111', 'end' => ['$exists' => false]], ['$set' => ['end' => $currentDate]], ['session' => $s1], ); $items->insertOne( ['sku' => '111-nuts', 'name' => 'Pecans', 'start' => $currentDate], ['session' => $s1], );
with client.start_session(causal_consistency=True) as s1: current_date = datetime.datetime.today() items = client.get_database( "test", read_concern=ReadConcern("majority"), write_concern=WriteConcern("majority", wtimeout=1000), ).items items.update_one( {"sku": "111", "end": None}, {"$set": {"end": current_date}}, session=s1 ) items.insert_one( {"sku": "nuts-111", "name": "Pecans", "start": current_date}, session=s1 )
let s1 = client1.startSession(options: ClientSessionOptions(causalConsistency: true)) let currentDate = Date() var dbOptions = MongoDatabaseOptions( readConcern: .majority, writeConcern: try .majority(wtimeoutMS: 1000) ) let items = client1.db("test", options: dbOptions).collection("items") let result1 = items.updateOne( filter: ["sku": "111", "end": .null], update: ["$set": ["end": .datetime(currentDate)]], session: s1 ).flatMap { _ in items.insertOne(["sku": "nuts-111", "name": "Pecans", "start": .datetime(currentDate)], session: s1) }
let s1 = client1.startSession(options: ClientSessionOptions(causalConsistency: true)) let currentDate = Date() var dbOptions = MongoDatabaseOptions( readConcern: .majority, writeConcern: try .majority(wtimeoutMS: 1000) ) let items = client1.db("test", options: dbOptions).collection("items") try items.updateOne( filter: ["sku": "111", "end": .null], update: ["$set": ["end": .datetime(currentDate)]], session: s1 ) try items.insertOne(["sku": "nuts-111", "name": "Pecans", "start": .datetime(currentDate)], session: s1)
If another client needs to read all current sku
values, you can
advance the cluster time and the operation time to that of the other
session to ensure that this client is causally consistent with the
other session and read after the two writes:
/* Make a new session, session2, and make it causally-consistent * with session1, so that session2 will read session1's writes. */ session2 = mongoc_client_start_session (client, session_opts, &error); if (!session2) { fprintf (stderr, "couldn't start session: %s\n", error.message); goto cleanup; } /* Set the cluster time for session2 to session1's cluster time */ cluster_time = mongoc_client_session_get_cluster_time (session1); mongoc_client_session_advance_cluster_time (session2, cluster_time); /* Set the operation time for session2 to session2's operation time */ mongoc_client_session_get_operation_time (session1, ×tamp, &increment); mongoc_client_session_advance_operation_time (session2, timestamp, increment); /* Run a find on session2, which should now find all writes done * inside of session1 */ find_opts = bson_new (); res = mongoc_client_session_append (session2, find_opts, &error); if (!res) { fprintf (stderr, "couldn't add session to opts: %s\n", error.message); goto cleanup; } find_query = BCON_NEW ("end", BCON_NULL); read_prefs = mongoc_read_prefs_new (MONGOC_READ_SECONDARY); cursor = mongoc_collection_find_with_opts (coll, query, find_opts, read_prefs); while (mongoc_cursor_next (cursor, &result)) { json = bson_as_relaxed_extended_json (result, NULL); fprintf (stdout, "Document: %s\n", json); bson_free (json); } if (mongoc_cursor_error (cursor, &error)) { fprintf (stderr, "cursor failure: %s\n", error.message); goto cleanup; }
using (var session2 = client.StartSession(new ClientSessionOptions { CausalConsistency = true })) { session2.AdvanceClusterTime(session1.ClusterTime); session2.AdvanceOperationTime(session1.OperationTime); var items = client.GetDatabase( "test", new MongoDatabaseSettings { ReadPreference = ReadPreference.Secondary, ReadConcern = ReadConcern.Majority, WriteConcern = new WriteConcern(WriteConcern.WMode.Majority, TimeSpan.FromMilliseconds(1000)) }) .GetCollection<BsonDocument>("items"); var filter = Builders<BsonDocument>.Filter.Eq("end", BsonNull.Value); foreach (var item in items.Find(session2, filter).ToEnumerable()) { // process item } }
// Example 2: Advance the cluster time and the operation time to that of the other session to ensure that // this client is causally consistent with the other session and read after the two writes. ClientSession session2 = client.startSession(ClientSessionOptions.builder().causallyConsistent(true).build()); session2.advanceClusterTime(session1.getClusterTime()); session2.advanceOperationTime(session1.getOperationTime()); items = client.getDatabase("test") .withReadPreference(ReadPreference.secondary()) .withReadConcern(ReadConcern.MAJORITY) .withWriteConcern(WriteConcern.MAJORITY.withWTimeout(1000, TimeUnit.MILLISECONDS)) .getCollection("items"); for (Document item: items.find(session2, eq("end", BsonNull.VALUE))) { System.out.println(item); }
async with await client.start_session(causal_consistency=True) as s2: s2.advance_cluster_time(s1.cluster_time) s2.advance_operation_time(s1.operation_time) items = client.get_database( "test", read_preference=ReadPreference.SECONDARY, read_concern=ReadConcern("majority"), write_concern=WriteConcern("majority", wtimeout=1000), ).items async for item in items.find({"end": None}, session=s2): print(item)
my $s2 = $conn->start_session({ causalConsistency => 1 }); $s2->advance_cluster_time( $s1->cluster_time ); $s2->advance_operation_time( $s1->operation_time ); $items = $conn->get_database( "test", { read_preference => 'secondary', read_concern => { level => 'majority' }, write_concern => { w => 'majority', wtimeout => 10000 }, } )->get_collection("items"); $cursor = $items->find( { end => undef }, { session => $s2 } ); for my $item ( $cursor->all ) { say join(" ", %$item); }
$s2 = $client->startSession( ['causalConsistency' => true], ); $s2->advanceClusterTime($s1->getClusterTime()); $s2->advanceOperationTime($s1->getOperationTime()); $items = $client->selectDatabase( 'test', [ 'readPreference' => new \MongoDB\Driver\ReadPreference(\MongoDB\Driver\ReadPreference::SECONDARY), 'readConcern' => new \MongoDB\Driver\ReadConcern(\MongoDB\Driver\ReadConcern::MAJORITY), 'writeConcern' => new \MongoDB\Driver\WriteConcern(\MongoDB\Driver\WriteConcern::MAJORITY, 1000), ], )->items; $result = $items->find( ['end' => ['$exists' => false]], ['session' => $s2], ); foreach ($result as $item) { var_dump($item); }
with client.start_session(causal_consistency=True) as s2: s2.advance_cluster_time(s1.cluster_time) s2.advance_operation_time(s1.operation_time) items = client.get_database( "test", read_preference=ReadPreference.SECONDARY, read_concern=ReadConcern("majority"), write_concern=WriteConcern("majority", wtimeout=1000), ).items for item in items.find({"end": None}, session=s2): print(item)
let options = ClientSessionOptions(causalConsistency: true) let result2: EventLoopFuture<Void> = client2.withSession(options: options) { s2 in // The cluster and operation times are guaranteed to be non-nil since we already used s1 for operations above. s2.advanceClusterTime(to: s1.clusterTime!) s2.advanceOperationTime(to: s1.operationTime!) dbOptions.readPreference = .secondary let items2 = client2.db("test", options: dbOptions).collection("items") return items2.find(["end": .null], session: s2).flatMap { cursor in cursor.forEach { item in print(item) } } }
try client2.withSession(options: ClientSessionOptions(causalConsistency: true)) { s2 in // The cluster and operation times are guaranteed to be non-nil since we already used s1 for operations above. s2.advanceClusterTime(to: s1.clusterTime!) s2.advanceOperationTime(to: s1.operationTime!) dbOptions.readPreference = .secondary let items2 = client2.db("test", options: dbOptions).collection("items") for item in try items2.find(["end": .null], session: s2) { print(item) } }
Limitations
The following operations that build in-memory structures are not causally consistent:
Operation | Notes |
---|---|
$collStats with latencyStats option. | |
Returns an error if the operation is associated with a causally
consistent client session. | |
Returns an error if the operation is associated with a causally
consistent client session. | |
Returns an error if the operation is associated with a causally
consistent client session. | |
Returns an error if the operation is associated with a causally
consistent client session. | |