[null,null,["最后更新时间 (UTC)：2025-08-29。"],[],[],null,["\u003cbr /\u003e\n\n\n|--------------------------------------------------------|\n| *Relevant to Cloud Firestore Enterprise edition only.* |\n\n\u003cbr /\u003e\n\nThis page describes the best practices when bulk loading data to Cloud Firestore with MongoDB compatibility\nwith tools like `mongoimport`.\n\nCloud Firestore is a highly distributed system offering automatic\nscaling to meet the needs of your business. Cloud Firestore dynamically\nsplits and combines your data based on the load received by the system.\n\nLoad-based splitting happens automatically without any required\npre-configuration. The Cloud Firestore load-based splitting system has some\nimportant, unique characteristics compared to other document databases that\nare important to keep in mind as you model your data.\n\nCloud Firestore's distributed nature can require changing some design choices\nto change, particularly for workloads that were optimized for\ndatabases where the primary replica is the bottleneck for write throughput.\n\nBest Practices\n\nWorkloads that process large amounts of data in a single threaded client can\ncreate a bottleneck. Clients might be able to use single threading to bulk load\ndata, as the throughput of the client and server are similarly matched.\nA Cloud Firestore database can handle significantly more parallelism, but\nthis requires that you configure clients to send requests in parallel.\n\n`mongoimport`\n\nWhen using the `mongoimport` tool, requests are made sequentially by default.\nTo improve the load time into Cloud Firestore,\nset the number of workers with the `--numInsertionWorkers` flag.\nThe correct setting might require tuning based on\nthe size of your client, but we generally recommend starting with at least `32`.\n\nasync programming\n\nWhen developing your own software using MongoDB compatible APIs, you can improve\nparallelism in the following ways:\n\n- *Async frameworks*: using async frameworks let you process and respond to requests in parallel. It is not necessary to develop any complex pooling or queuing when making calls to your database. Each request flow can use independent connections and make their database calls in parallel.\n- *Use parallelized compute offerings* : using services like Cloud Run, your system can scale the number of computation workers required to process data.\n\nTransient Failures\n\nWhen working with a large distributed system like Cloud Firestore, you might encounter\ntransient failures such as network blips or contention on a\ndocument.\n\nWhen bulk loading large amounts of information, it's important to\nmaintain a retry strategy for failed writes without failing the larger bulk load\noperation.\n| **Note:** Cloud Firestore with MongoDB compatibility does not support `retryWrites`. We recommend using transactions to ensure your application guarantees idempotency."]]
