mongoshake

• aliyun/mongo-shake的一个开源项目，一个通过拉取操作日志数据同步工具，但是无法双向同步 没有修改mongoDb内核的情况下，双向同步会循环写入
• 修改 部分代码在通过mongoshake写入时会向collection多插入一个字段(__go)以避免循环写入
• 修改 通过文件通道的逻辑，writer每隔一定时间将日志move到另一个目录，reader定时读取文件
• 对事物产生的日志(ns:admin$cmd,o:applyOps...)不进行过滤，接受后将applyOps拆分为多个独立操作

This is a brief introduction of Mongo-Shake, please visit english wiki or chinese wiki if you want to see more details including architecture, data flow, performance test, business showcase and so on.

• English document
• Chinese document
• FAQ document
• Performance test document

Mongo-Shake

Mongo-Shake is developed and maintained by DocumentDB Develop Team in Alibaba-Cloud.
Mongo-Shake is a universal platform for services based on MongoDB's oplog. It fetches oplog from source mongo database, and replays in the target mongo database or sends to other ends in different tunnels. If the target side is mongo database which means replay oplog directly, it's like a syncing tool that used to copy data from source MongoDB to another MongoDB to build redundant replication or active-active replication. Except for this direct way, there are others tunnel types such like rpc, file, tcp, kafka. Receivers wrote by users must define their own interfaces to connecting to these tunnels respectively. Users can also define there own tunnel type which is pluggable. If connecting to a third-party message middleware like kafka, the consumer can get the subscriber data in an asynchronous way in pub/sub module flexibly. Here comes general data flow,

The source can be either single mongod, replica set or sharding while target can be mongod or mongos. If the source is replica set, we suggest fetching data from secondary/hidden to ease the primary pressure. If the source is sharding, every shard should connect to Mongo-Shake. There can be several mongos on the target side to keep high availability, and different data will be hashed and written to different mongos.

Parallel Replication

There are three options for parallel replication which we call 'shad_key': id, collection and auto. id means the concurrency granularity is document while collection means the granularity is collection/table. auto option is decided by if there has unique index of any collection, it will change to collection with unique index exist otherwise id.

High Availability

Mongo-Shake periodically persistent its context into register center which by default is the source database. Currently, the context is checkpoint which marks the position of successfully replay oplog.
Hypervisor mechanism is also supported so that it will restart immediately when dies(master_quorum in configuration).

Filter

Support filtering database and collection namespace with whitelist and blacklist.

Global ID

In Aliyun internal version, global id(also called gid) is supported which marks the id of the database. It can be used to avoid loop when two databases become backup of each other. Mongo-Shake only fetches the oplogs equal to source database id, all the oplogs are be fetched when no gid gave. For current opensource version, it's not supported limited by the modification of MongoDB kernel. If you want to build active-active replication without gid supported, please visit FAQ document to see more details.

Tunnel

As mentioned above, we support several tunnel types such like: rcp, tcp, file, kafka, mock and direct. rpc and tcp means connecting to receiver synchronously by net/rcp and TCP respectively; file means writing output into file; kafka is an asynchronous way of sending the message; mock is used by testing that throws away all the data; direct means writing into target MongoDB directly. Users can also add or modify current tunnel type.
We offer receiver to connect to different tunnels like: rpc, tcp, file, mock and kafka. Please visit FAQ document to see more details.

Compressor

Gzip, zlib, deflate compressor are supported in batched oplogs before sending.

Monitor & Debug

User can monitor or debug Mongo-Shake through RESTful API, please visit FAQ document to see more details.

Other Details

Mongo-Shake uses mgo.v2 library which is a modification version of mgo.v2 to fetch oplogs from source MongoDB which is later than the given timestamp in configuration. Then, it filters oplogs based on whitelist, blacklist, and gid. All the oplogs will be transferred at least once which is acceptable because of idempotent of oplog DML. We use seq and ack to make sure the package is received which is similar to the sequence and acknowledgment numbers in TCP.
The oplogs are batched together in the handling pipeline.
Users can adjust the worker concurrency and executor concurrency according to the different environment.
Please see the detail documents listed at the beginning if you want to see more details.

code branch rules

version rules: a.b.c.

• a: major version
• b: minor version. even number means stable version.
• c: bugfix version

tag rules: add tag when releasing: "release-v{version}-{date}". for example: "release-v1.0.2-20180628"

usage

• git clone https://github.com/aliyun/mongo-shake.git
• cd mongo-shake/src/vendor
• GOPATH=`pwd`/../..; govendor sync #please note: must install govendor first and then pull all dependencies
• cd ../../ && ./build.sh
• ./bin/collector -conf=conf/collector.conf #please note: user must modify collector.conf first to match needs. You can also use "start.sh" script which supports hypervisor mechanism in Linux OS only.