diff --git a/docs/developer_guides/figs/guarantee-ts-consistency-relationship.png b/docs/developer_guides/figs/guarantee-ts-consistency-relationship.png
new file mode 100644
index 0000000000..1625a3bf03
Binary files /dev/null and b/docs/developer_guides/figs/guarantee-ts-consistency-relationship.png differ
diff --git a/docs/developer_guides/figs/guarantee-ts-do-search-right-now.png b/docs/developer_guides/figs/guarantee-ts-do-search-right-now.png
new file mode 100644
index 0000000000..269d46f0b1
Binary files /dev/null and b/docs/developer_guides/figs/guarantee-ts-do-search-right-now.png differ
diff --git a/docs/developer_guides/figs/guarantee-ts-ts-mask.png b/docs/developer_guides/figs/guarantee-ts-ts-mask.png
new file mode 100644
index 0000000000..6703d9cd50
Binary files /dev/null and b/docs/developer_guides/figs/guarantee-ts-ts-mask.png differ
diff --git a/docs/developer_guides/figs/guarantee-ts-wait-for-service-time.png b/docs/developer_guides/figs/guarantee-ts-wait-for-service-time.png
new file mode 100644
index 0000000000..27edffcbfa
Binary files /dev/null and b/docs/developer_guides/figs/guarantee-ts-wait-for-service-time.png differ
diff --git a/docs/developer_guides/how-guarantee-ts-works-cn.md b/docs/developer_guides/how-guarantee-ts-works-cn.md
new file mode 100644
index 0000000000..3b6a8a6660
--- /dev/null
+++ b/docs/developer_guides/how-guarantee-ts-works-cn.md
@@ -0,0 +1,62 @@
+# Guarantee Timestamp in Search Requests
+
+[English Version of This Doc](./how-guarantee-ts-works.md)
+
+很多同学接触 Milvus 时都会对 Search 请求里面茫茫多的参数感到迷惑不解，尤其是为 Milvus 开
+发 sdk 客户端的同学。这个文档就会介绍 Search 请求里面一个比较特殊的参数——“Guarantee
+Timestamp”，以下简称 “GuaranteeTs”。
+
+## Milvus 时钟机制
+
+像大多数分布式系统一样，Milvus 会为每一条进入系统的记录分配一个时间戳。与此同时，Milvus
+是一个存储计算分离的系统，数据持久化负载由 DataNodes 承担，最终会落盘到 MinIO/S3 之类的
+分布式对象存储中。Search 之类的计算任务由 QueryNodes 承担，计算链路会同时处理两类数据——
+批式数据以及流式数据。其中，批式数据不会再被更改，Search 请求会看到批式数据里面的所有数据。
+QueryNodes 和 DataNodes 通过同一个订阅机制消费用户的插入请求，这些构成了流式数据。由于
+存在网络延迟，QueryNodes 往往不会持有最新的流式数据。如果没有其他保障，在流式数据上直接做
+Search 会损失许多未消费的数据，降低查询精度，这个时候 GuaranteeTs 就派上用场了。
+
+Milvus 通过时间戳水印来保障读链路的一致性，如下图所示，在往消息队列插入数据时，
+Milvus 不光会为这些插入记录打上时间戳，还会不间断地插入同步时间戳，以图中同步时间戳
+syncTs1 为例，当下游消费者（比如QueryNodes）看到 syncTs1，那么意味着 syncTs1 以前的
+数据已经全部被消费了，换句话说，比 syncTs1 时间戳还小的插入记录不会再出现在消息队列中。
+当然了，有的话，那肯定是系统有 bug，如果你发现了，还希望尽快告诉我们。
+
+![ts-watermask](./figs/guarantee-ts-ts-mask.png)
+
+## Guarantee Timestamp
+
+上面说了，QueryNodes 会不断从消息队列里面拿到插入记录以及同步时间戳，每消费到一个同步时间
+戳，QueryNodes 会把这个时间戳称为可服务时间——“ServiceTime”，有了上图，ServiceTime
+实际上就很好理解了，意味着 QueryNodes 能够看到 ServiceTime 以前所有的数据了。
+
+有了这个 ServiceTime，Milvus 根据不同用户对一致性以及可用性的需求，提供了 GuaranteeTs，
+用户可以指定 GuaranteeTs 告知 QueryNodes 我这次 Search 请求必须看到 GuaranteeTs 以前
+的所有数据。
+
+如下图所示，如果 GuaranteeTs 小于 ServiceTime，QueryNodes 可以立刻执行 Search 查询。
+
+![do-search-right-now](./figs/guarantee-ts-do-search-right-now.png)
+
+如果 GuaranteeTs 大于 ServiceTime，QueryNodes 必须从消息队列里持续消费同步时间戳，
+直到 ServiceTime 大于 GuaranteeTs 才能执行 Search 查询。
+
+![wait-for-service-time](./figs/guarantee-ts-wait-for-service-time.png)
+
+如果用户希望得到足够高的查询精度，对一致性有较高的要求，对查询时延不敏感，那么 GuaranteeTs
+应该尽可能大；反之，如果用户希望尽快得到搜索结果，对可用性有较高的要求，对查询精度有较高的
+容忍程度，那么 GuaranteeTs 可以不必特别大。
+
+如下图所示，不同的 GuaranteeTs 分别对应四种不同的一致性：
+
+![relationship-between-consistency-and-guaranteeTs](./figs/guarantee-ts-consistency-relationship.png)
+
+- 强一致性：GuaranteeTs 设为系统最新时间戳，QueryNodes 需要等待 ServiceTime 推进到
+当前最新时间戳才能执行该 Search 请求；
+- 最终一致性：GuaranteeTs 设为 0，跳过一致性检查，立刻在当前已有数据上执行 Search 查询；
+- 有界一致性：GuaranteeTs 是一个比系统最新时间稍旧的时间，在可容忍范围内可以立刻执行查询；
+- 客户端一致性：客户端使用上一次写入的时间戳作为 GuaranteeTs，那么每个客户端至少能看到
+自己插入的全部数据。
+
+Milvus 默认提供强一致性，如果用户不传入 GuaranteeTs，那么会将 GuaranteeTs 设为系统
+当前的最新时间戳。
diff --git a/docs/developer_guides/how-guarantee-ts-works.md b/docs/developer_guides/how-guarantee-ts-works.md
new file mode 100644
index 0000000000..38a36fcc27
--- /dev/null
+++ b/docs/developer_guides/how-guarantee-ts-works.md
@@ -0,0 +1,80 @@
+# Guarantee Timestamp in Search Requests
+
+[中文文档](./how-guarantee-ts-works-cn.md)
+
+Users and developers may feel confused by the so many parameters in the search requests, especially for those developers
+who develop sdk client for Milvus.
+
+This document will explain a special parameter in the search request-"Guarantee Timestamp", hereinafter referred to as
+"GuaranteeTs".
+
+## Milvus Clock Mechanism
+
+Like most distributed systems, Milvus will assign a timestamp to each record that enters the system.
+
+At the same time, Milvus is a system that separates storage and computing. 
+
+The DataNodes handle the data persistence load, and the data will eventually be flushed to MinIO/S3 or other distributed
+object storage.
+
+QueryNodes handle the computing(read) requests. The reading link will process two types of data at the same time--batch 
+data and streaming data.
+
+Among them, the batch data will no longer be changed, and the search request will see all the data in the batch data.
+
+QueryNodes and DataNodes consume user insert requests through the same subscription mechanism, which constitute the
+streaming data.
+
+Due to the network delays, QueryNodes often do not hold all the latest streaming data. Without other control mechanism,
+some unconsumed data will be invisible and thus reduce the query accuracy. At this time, let's introduce the
+"GuaranteeTs".
+
+Milvus uses timestamp and timetick watermark to ensure the consistency of the read link.
+
+As shown in the figure below, when inserting data into the message queue, Milvus will not only time stamp these inserted
+records, but also insert timetick continuously. Taking "syncTs1" in the figure as an example, when downstream consumers
+(such as QueryNodes) see syncTs1, it means that all the data which is previous before syncTs1 has been consumed. In
+other words, the inserted record with a timestamp smaller than syncTs1 will no longer appear in the message queue.
+Of course, if there is, there must be a bug in the system. If you find it, we are pleased to solve it as soon as
+possible.
+
+![ts-watermask](./figs/guarantee-ts-ts-mask.png)
+
+## Guarantee Timestamp
+
+As mentioned above, QueryNodes will continue to get insert records and timetick from the message queue. Every time a
+timetick is consumed, QueryNodes will update this timetick as the serviceable time-"ServiceTime". With the above figure,
+it is easy to understand ServiceTime that QueryNodes can see all the data before ServiceTime.
+
+With this ServiceTime, Milvus provides GuaranteeTs according to the needs of different users for consistency and
+availability. Users can specify GuaranteeTs to inform QueryNodes that search request must see all the data before
+GuaranteeTs.
+
+As shown in the figure below, if GuaranteeTs is less than ServiceTime, QueryNodes can execute search request
+immediately.
+
+![do-search-right-now](./figs/guarantee-ts-do-search-right-now.png)
+
+If GuaranteeTs is greater than ServiceTime, QueryNodes must continue to consume timetick from the message queue. Search
+request cannot be executed until ServiceTime is greater than GuaranteeTs.
+
+![wait-for-service-time](./figs/guarantee-ts-wait-for-service-time.png)
+
+If users want to obtain high query accuracy, have higher requirements for consistency, and are not sensitive to query
+delay, then GuaranteeTs should be set to the latest wall clock; on the contrary, if users have higher requirements for
+availability, in other words, if they want to get the search results as soon as possible, then GuaranteeTs may not be
+set very large.
+
+As shown in the figure below, different GuaranteeTs correspond to four different consistency levels:
+
+![relationship-between-consistency-and-guaranteeTs](./figs/guarantee-ts-consistency-relationship.png)
+
+- Strong consistency: GuaranteeTs is set to the newest timestamp of the system, QueryNodes will wait for the ServiceTime
+to be greater than or equal to the GuaranteeTs;
+- Eventual consistency: Set GuaranteeTs to 0 and skip the check, QueryNodes will execute search requests immediately;
+- Bounded Staleness: Set GuaranteeTs to an older timestamp, such as 1 minutes ago, the query can be executed immediately
+within a tolerable range;
+- Read your own write (Session): Set GuaranteeTs to the client last write, in this way, every client will see all their
+own data.
+
+Milvus provides the strong consistency by default.