orderer

command
v0.0.0-...-e9fc09b Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Jan 8, 2020 License: Apache-2.0 Imports: 36 Imported by: 0

README

Hyperledger Fabric Ordering Service

The Hyperledger Fabric ordering service provides an atomic broadcast ordering service for consumption by the peers. This means that many clients can submit messages to the ordering service, and the same sequence of ordered batches will be delivered to all clients in response.

Protocol definition

The atomic broadcast ordering protocol for Hyperledger Fabric is described in hyperledger/fabric/protos/orderer/ab.proto. There are two services: the Broadcast service for injecting messages into the system and the Deliver service for receiving ordered batches from the service.

Service types

  • Solo ordering service (testing): The solo ordering service is intended to be an extremely easy to deploy, non-production ordering service. It consists of a single process which serves all clients, so consensus is not required as there is a single central authority. There is correspondingly no high availability or scalability. This makes solo ideal for development and testing, but not for deployment.
  • Kafka-based ordering service (production): The Kafka-based ordering service leverages the Kafka pub/sub system to perform the ordering, but wraps this in the familiar ab.proto definition so that the peer orderer client code does not to be written specifically for Kafka. Kafka is currently the preferred choice for production deployments which demand high throughput and high availability, but do not require byzantine fault tolerance.
  • PBFT ordering service (pending): The PBFT ordering service will use the Hyperledger Fabric PBFT implementation (currently under development) to order messages in a byzantine fault tolerant way.
Choosing a service type

In order to set a service type, the ordering service administrator needs to set the right value in the genesis block that the ordering service nodes will be bootstrapped from.

Specifically, the value corresponding to the ConsensusType key of the Values map of the Orderer config group on the system channel should be set to either solo or kafka.

For details on the configuration structure of channels, refer to the Channel Configuration guide.

configtxgen is a tool that allows for the creation of a genesis block using profiles, or grouped configuration parameters — refer to the Configuring using the connfigtxgen tool guide for more.

The location of this block can be set using the ORDERER_GENERAL_GENESISFILE environment variable. As is the case with all the configuration paths for Fabric binaries, this location is relative to the path set via the FABRIC_CFG_PATH environment variable.

Ledger types

Because the ordering service must allow clients to seek within the ordered batch stream, orderers need a backing ledger, where they maintain a local copy of past batches. Not all ledgers are crash fault tolerant, so care should be used when selecting a ledger for an application. Because the orderer ledger interface is abstracted, the ledger type for a particular orderer may be selected at runtime. The following options are available:

  • File ledger (production): The file-based ledger stores blocks directly on the file system. The block locations on disk are 'indexed' in a lightweight LevelDB database by number so that clients can efficiently retrieve a block by number. This is the default, and the suggested option for production deployments.
  • RAM ledger (testing): The RAM ledger implementation is a simple development oriented ledger which stores batches purely in memory, with a configurable history size for retention. This ledger is not crash fault tolerant; restarting the process will reset the ledger to the genesis block.
  • JSON ledger (testing): The file ledger implementation is a simple development oriented ledger which stores batches as JSON encoded files on the filesystem. This is intended to make inspecting the ledger easy and to allow for crash fault tolerance. This ledger is not intended to be performant, but is intended to be simple and easy to deploy and understand.
Choosing a ledger type

This can be set by setting the ORDERER_GENERAL_LEDGERTYPE environment variable before executing the orderer binary. Acceptable values are file (default), ram, and json.

Experimenting with the orderer service

To experiment with the orderer service you may build the orderer binary by simply typing go build in the hyperledger/fabric/orderer directory. You may then invoke the orderer binary with no parameters, or you can override the bind address, port, and backing ledger by setting the environment variables ORDERER_GENERAL_LISTENADDRESS, ORDERER_GENERAL_ LISTENPORT and ORDERER_GENERAL_LEDGER_TYPE respectively.

There are sample clients in the fabric/orderer/sample_clients directory.

  • The broadcast_timestamp client sends a message containing the timestamp to the Broadcast service.
  • The deliver_stdout client prints received batches to stdout from the Deliver interface.

These may both be built simply by typing go build in their respective directories. Note that neither of these clients supports config (so editing the source manually to adjust address and port is required), or signing (so they can only work against channels where no ACL is enforced).

Profiling

Profiling the ordering service is possible through a standard HTTP interface documented here. The profiling service can be configured using the orderer.yaml file, or through environment variables. To enable profiling set ORDERER_GENERAL_PROFILE_ENABLED=true, and optionally set ORDERER_GENERAL_PROFILE_ADDRESS to the desired network address for the profiling service. The default address is 0.0.0.0:6060 as in the Golang documentation.

Note that failures of the profiling service, either at startup or anytime during the run, will cause the overall orderer service to fail. Therefore it is currently not recommended to enable profiling in production settings.

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License. s

Documentation

The Go Gopher

There is no documentation for this package.

Directories

Path Synopsis
common
configupdate is an implementation of the broadcast.Proccessor interface It facilitates the preprocessing of CONFIG_UPDATE transactions which can generate either new CONFIG transactions or new channel creation ORDERER_TRANSACTION messages.
configupdate is an implementation of the broadcast.Proccessor interface It facilitates the preprocessing of CONFIG_UPDATE transactions which can generate either new CONFIG transactions or new channel creation ORDERER_TRANSACTION messages.
ram
mocks
sample_clients

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL