The only reason atomic modifies work against a single-collection is because the mongodb developers recently exchanged a database lock with a collection wide write-lock. Deciding that the increased concurrency here was worth the trade-off. At it's core, mongodb is a memory-mapped file: they've delegated the buffer-pool management to the machine's vm subsystem. Because it's always in memory, they're able to get away with very course grained locks: you'll be performing in-memory only operations while holding it, which will be extremely fast. This differs significantly from a traditional database system which is sometimes forced to perform I/O while holding a pagelock or a rowlock.

“在MongoDB中，单个文档上的操作是原子的”——这是过去的事情

在MongoDB 4.0的新版本中，您可以:

However, for situations that require atomicity for updates to multiple documents or consistency between reads to multiple documents, MongoDB provides the ability to perform multi-document transactions against replica sets. Multi-document transactions can be used across multiple operations, collections, databases, and documents. Multi-document transactions provide an “all-or-nothing” proposition. When a transaction commits, all data changes made in the transaction are saved. If any operation in the transaction fails, the transaction aborts and all data changes made in the transaction are discarded without ever becoming visible. Until a transaction commits, no write operations in the transaction are visible outside the transaction.

尽管对于如何执行和执行什么操作有一些限制。

检查Mongo Doc。 https://docs.mongodb.com/master/core/transactions/

The only reason atomic modifies work against a single-collection is because the mongodb developers recently exchanged a database lock with a collection wide write-lock. Deciding that the increased concurrency here was worth the trade-off. At it's core, mongodb is a memory-mapped file: they've delegated the buffer-pool management to the machine's vm subsystem. Because it's always in memory, they're able to get away with very course grained locks: you'll be performing in-memory only operations while holding it, which will be extremely fast. This differs significantly from a traditional database system which is sometimes forced to perform I/O while holding a pagelock or a rowlock.

在“星巴克不使用两阶段提交”中有一个很好的解释。

这不是关于NoSQL数据库，但它确实说明了一点，有时您可以承担丢失事务或使数据库暂时处于不一致的状态。

我不认为这是需要“修复”的东西。解决办法是使用acid兼容的关系数据库。当NoSQL的行为符合应用程序需求时，您可以选择一个NoSQL替代方案。

如果您的存储支持每个键的线性化和比较和设置(对于MongoDB是这样的)，您可以在客户端实现原子多键更新(可序列化的事务)。这种方法在谷歌的Percolator和CockroachDB中使用，但没有什么可以阻止你在MongoDB中使用它。

我已经创建了这种交易的一步一步的可视化。我希望它能帮助你理解它们。

如果您不介意读提交隔离级别，那么有必要看看Peter Bailis的RAMP事务。它们也可以在客户端为MongoDB实现。

我认为其他人已经给出了很好的答案。 然而，我想补充的是，有ACID NOSQL数据库(如http://ravendb.net/)。所以这不仅仅是NOSQL - no ACID vs关系型ACID....的决定