Skip to content 


Summaries are client-generated snapshots of the state of the document at a given sequence number. It consolidates the op log by serializing the data model.

Shape of a Summary

Summaries take the form of a tree consisting of blobs. Each layer in the tree is parallel to a part of the runtime model. The root node corresponds to the Container, and it contains the protocol information as well as any partially processed chunk op data.

graph LR A(["(root)"]) --> B([.protocol]) A --> S([_scheduler]) A --> C1([DataStore_1]) A --> C2([DataStore_2]) A --> CH([.chunks]) DS1 --> .DS([.fluid-object]) DS1 --> D1([DDS_1]) DS1 --> D2([DDS_2]) D1 --> .D([.attributes]) D1 --> SB(["(more blobs)"]) D1 --> ST(["(subtrees)"]) classDef tree fill:#faa,stroke:#000 class A,B,S,C1,C2,D1,D2,ST tree


The protocol subtree of the summary is generated by the server. The server takes responsibility of augmenting the proposed summary by tacking it on to the root of the tree while acknowledging it. The protocol contains information about the quorum: including client join/leave messages and quorum proposal data. It also contains the document attributes: document ID, sequence number, and minimum sequence number at that point in time.

graph LR A(["(root)"]) --> B([.protocol]) A --> S([_scheduler]) A --> C1([DataStore_1]) A --> C2([DataStore_2]) A --> CH([.chunks]) B --> QM([quorumMembers]) B --> QP([quorumProposals]) B --> QV([quorumValues]) B --> DA([attributes]) classDef tree fill:#faa,stroke:#000 class A,B,S,C1,C2 tree

Fluid Data Stores

The second layer consists of a subtree for each Fluid Data Store which has been attached and bound to the container. They must have a unique ID, which is generally decided by the runtime to prevent collisions. This will be a flattened list of all bound data stores. Nested data stores referenced by handles are irrelevant to the structure of this graph.

Each Fluid Data Store subtree will have a single metadata blob containing information about how to load its code.

Distributed data structures

The third layer consists of a subtree for every distributed data structure found within each Fluid Data Store. They also have a unique ID generated by the runtime. The parent node is the Fluid Data Store which the DDS is bound to, so this graph is irrelevant to how a DDS can reference another by its handle.

They will have a single metadata blob containing information for how to load the code given the registry collections found in higher layers. Each subtree may have other blobs or subtrees as needed.

Summary handles

Sometimes nodes/subtrees within the summary remain unchanged since the last successful summary, or want to be reused by the next summary. In this case, when uploading a summary, handles can be used in place of trees or blobs. Handles are pointers to nodes within the previous tree. The handle itself is just a full path to the node it is referencing, for example: “/dataStoreId/ddsId” would reference the subtree for the data structure with ID “ddsId” within the data store with ID “dataStoreId”. When uploading to storage, the driver uses this path in conjunction with the previously uploaded summary ID to resolve these handles.

Summary lifecycle

  1. Runtime generates summary tree (more details below).
  2. Runtime uploads summary tree to storage, which returns a handle.
  3. Runtime submits a “summarize” op to the server containing that uploaded summary handle.
  4. Ordering service on server stamps and broadcasts the “summarize” op.
  5. Another service on server responds to “summarize” op.
    • The server can reject the summary by sending a “summaryNack” op referencing the sequence number of the “summarize” op.
    • The server can accept the summary, but first it must serialize the protocol state and add it to the posted summary. Then it will need to send a “summaryAck” op with the new handle to the augmented summary.
  6. Runtime watches for “summaryAck”/“summaryNack” ops, using them as input to its heuristics determining when to generate summaries


The job of summarizing consists of multiple parts within the runtime. Currently all summarization work is done in a separate agent client, which helps by preventing it from having any local operations to deal with.

Summary manager

The summary manager runs on every connected client, and its primary role is to decide which client should be responsible for summarizing, and then spawning and managing the separate “summarizer” client.

A client is “elected” to be responsible for summarizing simply by being the oldest member of the quorum. The Summary manager is a state machine that checks for state changes any time it connects/disconnects or a member joins/leaves the quorum. Because client join/leave messages are sequenced in the quorum, it is sufficient to check if this client is first in the list of clients, since all clients will have this same information at any given sequence number. It’s important to keep this simple and deterministic to prevent accidentally spawning multiple summarizer clients or not spawning one at all.

Once a client decides it is responsible for spawning a summarizer client, it will make a request through the loader with special parameters to force a new container, and indicate that it is not interactive. It will call run on that spawned client and listen for it to disconnect/finish running.


Once a client spawns a summarizer client, the summarizer client will use the Summarizer object for heuristics around actually generating the summaries. The underlying data structure will watch the ops to keep track of what summaries have been seen so far, as well as how many ops/how much time has passed since the last summary.

The configuration for heuristics are provided by the server. They consist of several points:

  • maxOps – the maximum number of ops since the last successful summary before trying to summarize again
  • maxTime – the maximum amount of time since the last successful summary before trying to summarize again
  • idleTime – the amount of time to wait while idle before summarizing
  • maxAckWaitTimeout – the maximum amount of time to wait for an ack or nack from the server in response to a summary op

In general, the summarizer will wait for a break from ops before trying to summarize. This is the idleTime configuration. If the client receives no ops for idleTime (15 sec currently), then it will make a summary attempt. If an op comes in, it will reset the idle timer.

Even if ops are consistently coming in before the idleTime is hit, if maxOps (500 ops currently) ops are received or maxTime (1 minute currently) passes since the last successful summary, the client will attempt to summarize anyway.

The Summarizer defers to the rest of the runtime to actually generate the summary, upload it to storage, and submit the op. Once complete, it will watch for the summary op to be broadcast, and then keep waiting for the summary ack or nack op to come in from the server in response. It will not try to generate another summary while waiting for the ack/nack op. In most cases, the server should respond quickly with the ack/nack. In some bad cases, it may never come or take too long, in this case the client will not wait longer than maxAckWaitTimeout before trying to generate and send another summary.

Retry in “safe” mode

When an unexpected error is encountered while summarizing or the server sends a nack, the runtime will retry one more time in “safe” mode. This does two things:

  1. It first asks the server what the latest successful summary ID is. It does this to be resilient to issues with its own state tracking. This works around the issue where a summary is actually accepted by the server, but it fails to send an ack op.
  2. It then generates the summary with fullTree set to true. This will prevent any subtree reuse optimizations. It will try to generate the entire tree, regardless of whether it has changed since last summary or not.

Container runtime

The Container runtime is actually responsible for generating the summary, uploading it to storage, and submitting the op.

Generating the Summary

The Container runtime stores partially received chunked ops in a separate blob. It will then loop through all the Fluid Data Stores it has bound to it and call their respective summarize functions.

Each Fluid Data Store Context will store its metadata in a blob. It will then load/realize itself which loads its Fluid Data Store Runtime, and call its internal summarize function. The Fluid Data Store Runtime will then loop through each DDS it has and call their summarize functions.

Similarly, each DDS will store its metadata in a blob. It will ensure it is loaded and call its custom serialize function.

Summarizer Node

Within these layers of summary generation, common functionality is encapsulated by the Summarizer Node object. The Summarizer Nodes form a tree parallel to the tree of Container, Fluid Data Store Contexts, and Channel Contexts (distributed data structures). They are needed to track all the state of the completed and pending summaries relevant to each node. The state is needed in order to help each node decide whether it can be reused from the previous summary and whether it can generate a fallback “failure” summary in error scenarios.

They work by tracking the latest successful summary reference sequence number as well as the parent path and this node’s path part. The paths are important as they can actually change when a “failure” summary is generated. A “failure” summary is when a Fluid Data Store encounters an error while summarizing, if possible it will instead send a handle pointing to the last successful summary of this subtree and a blob of outstanding ops addressed to this data store since that summary. It does this to isolate the problem to a single data store. When this happens, the summary will have nested subtrees for every consecutive failure, which complicates the base path for child nodes in the tree.

They also decide when the subtrees can be reused by comparing their latest change sequence number with the last successful summary sequence number. Then they use their path tracking to figure out what path to send.