Skip to content

Latest commit

 

History

History
514 lines (420 loc) · 23.2 KB

adr-100-data-companion-push-api.md

File metadata and controls

514 lines (420 loc) · 23.2 KB

ADR 100: Data Companion Push API

Changelog

  • 2023-05-03: Mark as rejected in favour of ADR-101 (@thanethomson)
  • 2023-04-04: Update based on review feedback (@thanethomson)
  • 2023-02-28: Renumber from 082 to 100 (@thanethomson)
  • 2022-01-05: Sync requirements with those of ADR 101 (@thanethomson)
  • 2022-12-18: Renamed proposal to "Data Companion Push API" (@thanethomson)
  • 2022-11-26: Clarify user stories and alternatives, allow for selective publishing of data via the companion API, buffer on disk instead of in memory (@thanethomson)
  • 2022-09-10: First draft (@thanethomson)

Status

Rejected

Context

At present, CometBFT handles a mixture of application- and consensus-related concerns, which results in code bloat and scope creep into territory that should technically be taken care of by the application. Some major examples of this right now are event and block execution result data, with CometBFT being responsible for:

  1. Storing event/execution result data.
    1. For some applications, this results in dramatic storage growth, leading to high operational costs for operators, as application-specific pruning is not feasible to implement within the consensus engine.
    2. Storing application-specific data also means that multiple underlying databases need to be supported (via cometbft-db) in order to cater to different applications' needs, which substantially increases the project's maintenance costs and complexity.
  2. Facilitating complex querying of this data, including providing real-time notifications of complex updates.
    1. This has resulted in needing to provide query mechanisms such as /block_search and /tx_search, which are not really optimal to provide via some of the databases supported by cometbft-db, and therefore result in problems like #517.
    2. This has also resulted in having to maintain multiple indexers and our own custom, database-independent indexing scheme, and having to turn down users' requests for even more indexers that would work better for their use cases, since this imposes a substantial maintenance burden on the core team.
    3. The current (somewhat unreliable) event subscription implementation can cause back-pressure into consensus and affect IBC relayer stability (see #6729 and #7156)

Ultimately the core team, whose focus is providing a high quality consensus engine, has finite resources. Having to take care of both consensus- and application-related concerns means that neither will be taken care of effectively.

It then starts to become clearer that, in order to address these problems in the long run, event and execution results data need to be handled in application territory. But what does the trajectory look like to facilitate this evolution over time?

Roadmap

This ADR proposes the following path to eventually separating these concerns:

  1. Provide an API explicitly dedicated to facilitating offloading of this data to a single data companion ("sidecar") service - this is the focus of this ADR. The primary function of such a companion is effectively to act as an ingest of sorts, translating block and block result data (including events) into application-specific data. Single is emphasized here, as we anticipate that attaching more than one companion service to a node may slow it down and affect its operations.
  2. Provide a reference implementation of a data companion which facilitates offering most of the current RPC endpoints from a service external to the node. Beyond serving as an example for the community as to how to implement a data companion, this will potentially:
    1. Allow for horizontal scalability of the RPC independently of the associated full node, reducing operational costs for operators who want to provide RPC services for their user base (who, in some cases, have to run multiple full nodes in order to facilitate a highly available RPC service).
    2. Address some of the problems outlined in RFC-006 relating to event subscription.
    3. Address more general problems in the RPC, such as its inherent susceptibility to DDoS attacks.
  3. Once the community has mostly migrated to using this data companion API, we can mark many of the node-based RPC endpoints as deprecated, and eventually remove them from the node.
  4. In parallel, start an effort to design a future where event and execution result data are handled exclusively in the application domain. Once these changes are ready to be rolled out, the ingest components of data companions are the only parts of the ecosystem that will need to be migrated to interact directly with the application instead of the node (e.g. by way of RPC endpoints exposed by the application, instead of the consensus engine).
  5. Once the ingests are migrated to rely on the application, it will be safe to entirely remove any notion of event data storage and retrieval, as well as indexing from the consensus engine, dramatically simplifying the consensus engine.

Alternative Approaches

  1. One clear alternative to this would be the approach outlined in ADR 075 (specifically, implementing it in isolation). This approach:

    1. Still leaves CometBFT responsible for maintaining a query interface and event indexing functionality, increasing the long-term maintenance burden of, and the possibility of feature sprawl in, that subsystem. To overcome this, we could remove the query interface altogether and just always publish all data.
    2. Only keeps track of a sliding window of events (and does so in-memory), and so does not provide reliable guaranteed delivery of this data. Even just persisting this data will not solve this problem: if we prioritize continued operation of the node over guaranteed delivery of this data, we may eventually run out of disk storage. So here we would need to persist the data and we would need to panic if the sliding window of events fills up.
    3. Allows for multiple consumers. While desirable from typical web services, CometBFT is a consensus engine and, as such, we should prioritize consensus over providing similar guarantees to a scalable RPC interface (especially since we know this does not scale, and if such a job could just be outsourced). If we have a single consumer, we can easily automatically prune data that was already sent. We could overcome this by limiting the API to a single consumer and single event window.

    Making the requisite modifications to ADR-075 basically converges on the solution outlined in this ADR, except that ADR-075 publishes data via JSON over HTTP while this solution proposes gRPC.

    Technically, however, ADR-075 is orthogonal to the solution proposed in this ADR, and could be implemented as part of an "RPC companion" to provide better delivery guarantees to subscribers.

  2. Another alternative is proposed in ADR 101, which has the same requirements as this proposed approach, except it implements a "pull" model instead.

  3. We could pick a database that provides real-time notifications of inserts to consumers and just dump the requisite data into that database. Consumers would then need to connect to that database, listen for updates, and then transform/process the data as needed.

    This approach could also inevitably lead to feature sprawl in people wanting us to support multiple different databases. It is also not clear which database would be the best option here, as that would most likely be influenced by integrators' use cases.

    Beyond this, if we exclusively support one database but integrators needed another for their use case, they would be forced to perform some kind of continuous ETL operation to transfer the data from the one database to the other, potentially doubling their storage capacity requirements and therefore their storage-related costs. This would also increase pruning complexity.

Decision

TODO(thane)

Detailed Design

The model proposed in this ADR is one where the node is a client, and the data companion is a server. This model is inverted in ADR-101 (the "pull" companion). The model in this ADR provides strong guarantees as to reliable delivery, but also potentially facilitates halting of consensus if the interaction with the companion is unreliable.

The model proposed in ADR-101 offers much weaker delivery guarantees, allowing the companion to lag significantly behind consensus, but offers the greatest potential uptime for the consensus engine.

Use Cases

  1. Integrators could provide standalone RPC services that would be able to scale independently of the CometBFT node. They would also be able to present the data in whatever format and by whichever protocol they want.
  2. Block explorers could make use of this data to provide real-time information about the blockchain.
  3. IBC relayer nodes could use this data to filter and store only the data they need to facilitate relaying without putting additional pressure on the consensus engine (until such time that a decision is made on whether to continue providing event data from CometBFT).

Requirements

  1. A node must support at most one data companion.

  2. All or part of the following data must be obtainable by the companion:

    1. Committed block data
    2. Block execution results (including event data) for blocks that have been committed
  3. The companion must be able to establish the earliest height for which the node has all of the requisite data.

  4. The API must be (or be able to be) appropriately shielded from untrusted consumers and abuse. Note that the companion is considered a trusted consumer of this API. Critical control facets of the API (e.g. those that influence the node's pruning mechanisms) must be implemented in such a way as to eliminate the possibility of accidentally exposing those endpoints to the public internet unprotected.

  5. The node must know, by way of signals from the companion, which heights' associated data are safe to automatically prune.

  6. The API must be opt-in. When off or not in use, it should have no impact on system performance.

  7. It must not slow down consensus operations without halting the node.

  8. It must not cause unbounded memory growth.

  9. It must provide one or more ways for operators to control storage growth.

  10. It must provide insight to operators (e.g. by way of logs/metrics) to assist in dealing with possible failure modes.

Entity Relationships

The following simple diagram shows the proposed relationships between CometBFT, a socket-based ABCI application, and the proposed data companion service.

flowchart RL
    comet[CometBFT]
    companion[Data Companion]
    app[ABCI App]

    comet --> app
    comet --> companion
Loading

As can be seen in this diagram, CometBFT (client) connects out to both the ABCI application (server) and data companion service (server) based on the CometBFT node's configuration.

The fact that CometBFT connects out to the companion service instead of the other way around provides a natural constraint on the number of consumers of the API. This also implies that the data companion is a trusted external service, just like an ABCI application, as it would need to be configured within the CometBFT configuration file.

gRPC API

The following gRPC API is proposed:

import "tendermint/abci/types.proto";
import "tendermint/types/block.proto";

// DataCompanionService allows CometBFT to publish certain data generated by
// the consensus engine to a single external consumer with specific reliability
// guarantees.
//
// Note that implementers of this service must take into account the possibility
// that CometBFT may re-send data that was previously sent. Therefore
// the service should simply ignore previously seen data instead of responding
// with errors to ensure correct functioning of the node, as responding with an
// error is taken as a signal to CometBFT that it must crash and await
// operator intervention.
service DataCompanionService {
    // BlockCommitted is called after a block has been committed. This method is
    // also called on CometBFT startup to ensure that the service received the
    // last committed block, in case there was previously a transport failure.
    //
    // If an error is returned, CometBFT will crash.
    rpc BlockCommitted(BlockCommittedRequest) returns (BlockCommittedResponse) {}
}

// BlockCommittedRequest contains at least the data for the block that was just
// committed. If enabled, it also contains the ABCI FinalizeBlock response data
// related to the block in this request.
message BlockCommittedRequest {
    // The block data for the block that was just committed.
    tendermint.types.Block                         block = 1;
    // The FinalizeBlockResponse related to the block in this request. This
    // field is optional, depending on the CometBFT node's configuration.
    optional tendermint.abci.FinalizeBlockResponse finalize_block_response = 2;
}

// BlockCommittedResponse is either empty upon succes, or returns one or more
// errors. Note that returning any errors here will cause CometBFT to crash.
message BlockCommittedResponse {
    // If an error occurred during the companion's processing of the request.
    optional Error error = 1;
}

// Error can be one of several different types of errors.
message Error {
    oneof error {
        UnexpectedBlockError unexpected_block_err = 1;
        UnexpectedFieldError unexpected_field_err = 2;
        ExpectedFieldError   expected_field_err   = 3;
    }
}

// UnexpectedBlockError is returned by the server when CometBFT sent it a
// block that is ahead of the block expected by the server.
message UnexpectedBlockError {
    // The height of the block expected by the server.
    int64 expected_height = 1;
}

// UnexpectedFieldError is returned by the server when CometBFT sent it a
// message containing an unexpected field. For instance, if the companion
// expects only the `block` field but not `finalize_block_response`, but somehow
// the companion receives a value for `finalize_block_response`, it should
// return this error because it indicates that either CometBFT or the data
// companion are most likely incorrectly configured.
message UnexpectedFieldError {
    // The name of the field whose value was expected to be empty, but was not.
    string field_name = 1;
}

// ExpectedFieldError is returned by the server when CometBFT sent it a
// message containing an empty field value for a field it was expecting to be
// populated. If this occurs, it indicates that either CometBFT or the data
// companion are most likely incorrectly configured.
message ExpectedFieldError {
    // The name of the field whose value was expected to be populated, but was
    // empty.
    string field_name = 1;
}

Request Buffering

In order to ensure reliable delivery, while also catering for intermittent faults, we should facilitate buffering of data destined for the companion service. This data would need to be stored on disk to ensure that CometBFT could attempt to resubmit all unsubmitted data upon CometBFT startup.

At present, this data is already stored on disk, and so practically we would need to implement some form of background pruning mechanism to remove the data we know has been shared with the companion service.

In order to not allow disk storage to grow unbounded, and ensure that we do not lose any critical data, we need to impose some kind of limit on the number of requests we will buffer before crashing the CometBFT node. In case there is manual intervention needed to restore the companion service, CometBFT should attempt to flush this buffer upon startup before continuing normal operations.

Crash Recovery

In order to ensure that the companion has the latest data, if CometBFT does not have confirmation that the companion has received the data for a particular height, it should resend all data from the last height for which it has confirmation.

An example where the companion saves all data up to and including height 10, but crashes before confirming receipt of data for height 11:

sequenceDiagram
    participant N as Node
    participant C as Companion

    N->>+C: BlockCommittedRequest (height = 10)
    C->>-N: BlockCommittedResponse (error = nil)
    N->>N: Save prune height as 10

    N->>C: BlockCommittedRequest (height = 11)
    break Node or companion crashes
        N-->C: Node and companion recover
    end

    N->>+C: BlockCommittedRequest (height = 11)
    C->>-N: BlockCommittedResponse (error = nil)
    N->>N: Save prune height as 11
Loading

In such a case, it is possible for the companion to receive the same height's data multiple times (at-least-once delivery). It is up to the companion to handle this correctly.

Configuration

The following configuration file update is proposed to support the data companion API.

[data_companion]
# By default, the data companion service interaction is disabled. It is
# recommended that this only be enabled on full nodes and not validators so as
# to minimize the possibility of network instability.
enabled = false

# Address at which the gRPC companion service server is hosted. It is
# recommended that this companion service be co-located at least within the same
# data center as the CometBFT node to reduce the risk of network latencies
# interfering in node operation.
addr = "http://localhost:26659"

# Controls the BlockCommitted gRPC call.
[data_companion.block_committed]
# Enable the BlockCommitted gRPC call. Only relevant if the data companion is
# enabled.
enabled = true
# Additional fields to publish in each BlockCommittedRequest sent to the
# companion. Available options:
# - "finalize_block_response": Also publish the FinalizeBlock response related
#                              to the block in the BlockCommittedRequest.
additionally_publish = ["finalize_block_response"]

Monitoring

To monitor the health of the interaction between CometBFT and the companion service, the following additional Prometheus metrics are proposed:

  • data_companion_send_time - A gauge that indicates the maximum time, in milliseconds, taken to send a single request to the companion service take from a rolling window of tracked send times (e.g. the maximum send time over the past minute).
  • data_companion_send_bytes - A counter indicating the number of bytes sent to the data companion since node startup.
  • data_companion_buffer_size - A gauge indicating the number of requests currently being buffered on disk on behalf of the companion service for each kind of request.
  • data_companion_buffer_size_bytes - A gauge indicating how much data is being buffered by CometBFT on disk, in bytes, on behalf of the companion service, for each kind of request.

Implications

  1. We will be able to mark the following RPC APIs as deprecated and schedule them for removal in a future release:

    1. The WebSocket subscription API
    2. /tx_search
    3. /block_search
    4. /block_results

    We may eventually be able to remove the /broadcast_tx_commit endpoint, but this is often useful during debugging and development, so we may want to keep it around.

  2. We will be able to remove all event indexing from CometBFT once we remove the above APIs.

  3. Depending on the implementation approach chosen, we will still need to store some quantity of data not critical to consensus. This data can automatically be pruned once CometBFT has successfully transmitted it to the companion service.

Release Strategy

As captured in the requirements, we should be able to release this as an additive, opt-in change, thereby not impacting existing APIs. This way we can evolve this API until such time that consumers are satisfied with removal of the old APIs mentioned in implications.

Follow-Up Work

If implemented, we should consider releasing our own RPC companion server with easy deployment instructions for people to test it out and compare the performance with directly connecting to the CometBFT node.

This RPC companion server could easily implement an interface like that outlined in ADR-075, providing the same HTTP- and cursor-based subscription mechanism, but storing data to and pulling data from a different database (e.g. PostgreSQL).

Consequences

Positive

  • Integrators can expose CometBFT data in whatever format/protocol they want (e.g. JSON-RPC or gRPC).
  • Integrators can index data in whatever way suits them best.
  • All users eventually benefit from faster changes to CometBFT because the core team has less, and less complex, code to maintain (implementing this ADR would add more code in the short-term, but would pave the way for significant reductions in complexity in future).
  • Can be rolled out as experimental and opt-in in a non-breaking way.
  • The broad nature of what the API publishes lends itself to reasonable long-term stability.

Negative

  • Keeping existing CometBFT functionality would involve operators having to run an additional process, which increases operational complexity.
  • It is unclear at present as to the impact of the requirement to publish large quantities of block/result data on the speed of block execution. This should be quantified in production networks as soon as this feature can be rolled out as experimental. If the impact is meaningful, we should either remove the feature or develop mitigation strategies (e.g. allowing for queries to be specified via the configuration file, or supporting a specific subset of use cases' data).
  • Requires a reasonable amount of coordination work across a number of stakeholders across the ecosystem in order to ensure their use cases are addressed effectively and people have enough opportunity to migrate.

Neutral

References