milvus-backup 0.6 快照备份格式让恢复提速约 36%
Milvus 备份恢复要重建索引很慢，milvus-backup 0.6 用快照格式连索引一起还原，测试快了 36%，做迁移或容灾演练的可以直接试。
Zilliz 的 milvus-backup 0.6 在 Milvus 3.0.1 及以上版本采用快照备份格式，恢复时无需逐行导入扫描，索引也会随数据一并还原。官方测试使用 3 个集合、每个 1000 万条 768 维向量，共约 92 GB 数据和 32 GB 索引，快照恢复用时 324 秒且包含索引，而旧的 binlog 格式恢复需 505 秒且索引仍要重建，节省约 36% 时间。恢复完成后加载集合并验证查询即可。
Rebuilding vector indexes can add a long wait when restoring a Milvus backup. On Milvus 3.0.1 and later, 𝗺𝗶𝗹𝘃𝘂𝘀-𝗯𝗮𝗰𝗸𝘂𝗽 𝟬.𝟲 𝘂𝘀𝗲𝘀 𝘁𝗵𝗲 𝘀𝗻𝗮𝗽𝘀𝗵𝗼𝘁 𝗯𝗮𝗰𝗸𝘂𝗽 𝗳𝗼𝗿𝗺𝗮𝘁 to restore data and existing indexes together. Snapshot restores avoid the row-by-row import scan and index rebuild required by the older binlog backup format. After restoration, you can load the collection and verify your queries. In our internal test, we used three collections with 10 million 768-dimensional vectors each: roughly 92 GB of data and 32 GB of indexes. • Snapshot restore: 𝟯𝟮𝟰 𝘀𝗲𝗰𝗼𝗻𝗱𝘀, including indexes. • Binlog restore: 𝟱𝟬𝟱 𝘀𝗲𝗰𝗼𝗻𝗱𝘀, with indexes still to rebuild. That’s about 𝟯𝟲% 𝗹𝗲𝘀𝘀 𝘁𝗶𝗺𝗲 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗿𝗲𝘀𝘁𝗼𝗿𝗲 𝘁𝗮𝘀𝗸 in this test. Collection loading and query validation come afterward. For your next migration or recovery drill, create a snapshot backup so you can github.com/zilliztech/mil… ng with the data. Learn how to restore Milvus from a snaphot: https://t.co/szRi6ALNT1 💬 0 🔄 0 ❤️ 0 👀 58 ⚡
