Skip to content
This repository has been archived by the owner on Apr 10, 2024. It is now read-only.
/ moresql Public archive

MoreSQL streams changes from MongoDB to Postgres in realtime

Notifications You must be signed in to change notification settings

zph/moresql

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

55 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

MoreSQL

[](NOTE: README.md is a generated FILE changes belong in docs/README.template.md. Update with make docs) Build Status GoDoc

Introduction

MoreSQL streams changes occuring in Mongo database into a Postgres db. MoreSQL tails the oplog and generates appropriate actions against Postgres. MoreSQL has the ability to do full synchronizations using UPSERTS, with the benefit over INSERTS that this can be executed against tables with existing data.

MoreSQL gives you a chance to use more sql and less mongo query language.

Project maturity: Moresql has 3 years of running in production for 2+ companies on their business critical workloads for core systems. It is a stable project with low maintenance and high performance.

QuickStart

Introduction

  • Create metadata table
  • Setup moresql.json (see Configuration)
  • Setup any recipient tables in postgres
    • Validate with ./moresql -validate
  • Deploy moresql binary to server from Github Releases
  • Configure Environmental variables
  • Run ./moresql -tail to start transmitting novelty
  • Run ./moresql -full-sync to populate the database
  • Write more sql ;D

Usage

Basic Use

Configuration

moresql.json configuration structure

{
   "DB_NAME": {
      "collections": {
         "COLLECTION_NAME": {
            "name": "COLLECTION_NAME",
            "pg_table": "PG_TABLE_NAME",
            "fields": {
               ...
            }
         },
         "COLLECTION_NAME2": {
            "name": "COLLECTION_NAME2",
            "pg_table": "PG_TABLE_NAME2",
            "fields": {
               ...
            }
         }
      }
   }
}

Field attributes have a simple and complex format.

The simple format is where you want to use the mongo field name as the postgres column name. In this format fields take the mongo field name (as key) and its postgres type (as value). The value can be any valid postgres type (https://www.postgresql.org/docs/current/datatype.html)

            "fields": {
               "_id": "TEXT",
               "addresses": "JSONB",
            }

The complex format allows for renaming fields, extracting nested keys and some advanced operations. See gjson project for full syntax details of dot notation: https://github.com/tidwall/gjson#path-syntax

            "fields": {
               "_id": "TEXT",
               // preferences is an field name that has an object in it. The object has a key of unsubscribe.
               // The following will promote this nested unsubscribe to a top level value as a postgres column
               "preferences.unsubscribe": {
                  "Postgres": {"Name": "is_unsubscribed", "Type": "BOOLEAN"},
                  "Mongo": {"Name": "preferences.unsubscribe", "Type": "object"}
               },
               // This extraction will look in books field with [{product_id: 1}, {product_id: 2}]
               // And for each object in there, fetch the product_id and combine it into an array of product ids [1, 2]
               // That value will be inserted into a postgres column called `book_ids`.
               "books.#.product_id": {
                  "Postgres": {
                  "Name": "book_ids",
                  "Type": "JSONB"
                  },
                  "Mongo": {
                  "Name": "books.#.product_id",
                  "Type": "object"
                  }
               },
            }

See examples/moresql.json for a full configuration

Tail

./moresql -tail -config-file=moresql.json

Tail is the primary run mode for MoreSQL. When tailing, the oplog is observed for novely and each INSERT/UPDATE/DELETE is translated to its SQL equivalent, then executed against Postgres.

Tail makes a best faith effort to do this and uses checkpoint markers to track last successfully applied Mongo Oplog event.

Given that tail mode executes UPSERTS instead of INSERT || UPDATE, we expect MoreSQL to be roughly eventually consistent. We're chosing to prioritize speed of execution (multiple workers) in lieu of some consistency. This helps to keep low latency with larger workloads. We currently partition workload among multiple workers but ensure that each collection.id combination will be routed to same worker in correct oplog order. This avoids the circumstance where two operations against same collection.id are executed by different workers, out of order.

Full Sync

./moresql -full-sync -config-file=moresql.json

Full sync is useful when first setting up a MoreSQL installation to port the existing Mongo data to Postgres. We recommend setting up a tailing instance first. Once that's running, do a full sync in different process. This should put the Mongo and Postgres into identical states.

Given the nature of streaming replica data from Mongo -> Postgres, it's recommended to run full sync at intervals in order to offset losses that may have occured during network issues, system downtime, etc.

Documentation

https://zph.github.io/moresql/

GoDoc

Table Setup

-- Execute the following SQL to setup table in Postgres. Replace $USERNAME with the moresql user.
-- create the moresql_metadata table for checkpoint persistance
CREATE TABLE public.moresql_metadata
(
    app_name TEXT NOT NULL,
    last_epoch INT NOT NULL,
    processed_at TIMESTAMP WITH TIME ZONE DEFAULT NOW() NOT NULL
);
-- Setup mandatory unique index
CREATE UNIQUE INDEX moresql_metadata_app_name_uindex ON public.moresql_metadata (app_name);

-- Grant permissions to this user, replace $USERNAME with moresql's user
GRANT SELECT, UPDATE, DELETE ON TABLE public.moresql_metadata TO $USERNAME;

COMMENT ON COLUMN public.moresql_metadata.app_name IS 'Name of application. Used for circumstances where multiple apps stream to same PG instance.';
COMMENT ON COLUMN public.moresql_metadata.last_epoch IS 'Most recent epoch processed from Mongo';
COMMENT ON COLUMN public.moresql_metadata.processed_at IS 'Timestamp for when the last epoch was processed at';
COMMENT ON TABLE public.moresql_metadata IS 'Stores checkpoint data for MoreSQL (mongo->pg) streaming';

Building Binary

Compile binary using make build

Commandline Arguments / Usage

Execute ./moresql --help

./bin/moresql
Repo https://github.com/zph/moresql
Usage of ./bin/moresql:
  -allow-deletes
     Allow deletes to propagate from Mongo -> PG (default true)
  -app-name string
     AppName used in Checkpoint table (default "moresql")
  -checkpoint
     Store and restore from checkpoints in PG table: moresql_metadata
  -config-file string
     Configuration file to use (default "moresql.json")
  -create-table-sql
     Print out the necessary SQL for creating metadata table required for checkpointing
  -enable-monitor
     Run expvarmon endpoint
  -error-reporting string
     Error reporting tool to use (currently only supporting Rollbar)
  -full-sync
     Run full sync for each db.collection in config
  -memprofile string
     Profile memory usage. Supply filename for output of memory usage
  -mongo-url MONGO_URL
     MONGO_URL aka connection string
  -postgres-url POSTGRES_URL
     POSTGRES_URL aka connection string
  -replay-duration duration
     Last x to replay ie '1s', '5m', etc as parsed by Time.ParseDuration. Will be subtracted from time.Now()
  -replay-second int
     Replay a specific epoch second of the oplog and forward from there.
  -ssl-cert string
     SSL PEM cert for Mongodb
  -ssl-insecure-skip-verify
     Skip verification of Mongo SSL certificate ala sslAllowInvalidCertificates
  -tail
     Tail mongodb for each db.collection in config
  -validate
     Validate the postgres table structures and exit

Validation of Configuration + Postgres Schema

./moresql -validate

This will report any issues related to the postgres schema being a mis-match for the fields and tables setup in configuration.

Requirements, Stability and Versioning

MoreSQL is expected and built with Golang 1.6, 1.7 and master in mind. Broken tests on these versions indicates a bug.

MoreSQL requires Postgres 9.5+ due to usage of UPSERTs. Using UPSERTs simplifies internal logic but also depends on UNIQUE indexes existing on each _id column in Postgres. See moresql -validate for advice.

Miscellanea

Error Reporting

Available through Rollbar. PRs welcome for other services. We currently use Rollus which reports errors synchronously. If this is a performance bottleneck please PR or issue.

Enable this by two steps:

export ERROR_REPORTING_TOKEN=asdfasdfasdf
export APP_ENV=[production, development, or staging]

And when running application use the following flag to enable reporting:

./moresql -tail -error-reporting "rollbar"

If these steps are not followed, errors will be reported out solely via logging.

Environmental Variables used in Moresql

MONGO_URL
POSTGRES_URL
ERROR_REPORTING_TOKEN
APP_ENV
DYNO
LOG_LEVEL

Mongo types

We guard against a few of these for conversion into Postgres friendly types.

Objects and Arrays do not behave properly when inserting into Postgres. These will be automatically converted into their JSON representation before inserting into Postgres.

As of writing, any BsonID/ObjectId should be noted as id type in Fields.Mongo.Type to facilitate this. In the future we may assume that all fields ending in _id are Id based fields and require conversion.

Converting from MoSQL

Read through configuration structure and convert your collections.yml into the new json format.

Unsupported Features

These features are part of mosql but not implemented in MoreSQL. PRs welcome.

  • extra_props field for spare data
  • Automatic creation of tables/columns (we require explicit actions from users after providing guidance)

Dot notation

We support dot notation for extracting nested objects from Mongo.

user.address will perform a (get_in map [:user :address]) type nested fetch.

See gjson project for full syntax details: https://github.com/tidwall/gjson#path-syntax

Performance

During benchmarking when moresql is asked to replay existing events from oplog we've seen the following performance with the following configurations:

5 workers per collection 500 generic workers On a Heroku 1X dyno

~ $ ./moresql -tail -replay-duration "5000m" | grep "Rate of"
{"level":"info","msg":"Rate of insert per min: 532","time":"2017-02-23T01:49:31Z"}
{"level":"info","msg":"Rate of update per min: 44089","time":"2017-02-23T01:49:31Z"}
{"level":"info","msg":"Rate of delete per min: 1","time":"2017-02-23T01:49:31Z"}
{"level":"info","msg":"Rate of read per min: 91209","time":"2017-02-23T01:49:31Z"}
{"level":"info","msg":"Rate of skipped per min: 46587","time":"2017-02-23T01:49:31Z"}

Approximately 700 updates/sec and 1500 reads/sec is our top observed throughput so far. Please submit PRs with further numbers using a similar command.

We expect the following bottlenecks: connection count in Postgres, pg connection limitations in Moresql (for safety), network bandwidth, worker availability.

At this level of throughput, Moresql uses ~90MB RAM. At low idle throughput of 10-20 req/sec it consumes ~30MB RAM.

In another benchmark when updating 28k documents simultaneously, we observed mean lag of ~ 500ms and 95% of requests arrived in <= 1194ms between when the document was updated in Mongo and when it arrived in Postgres.

See full performance information

For a general discussion of UPSERT performance in Postgres: https://mark.zealey.org/2016/01/08/how-we-tweaked-postgres-upsert-performance-to-be-2-3-faster-than-mongodb

Binaries

We release binaries for semvar tags on Github Releases page using goreleaser for the platforms listed in goreleaser.yml.

Credit and Prior Art

  • MoSQL - the project we used for 3 yrs at work and then retired with MoreSQL. Thanks Stripe!
  • GTM - the go library that builds on mgo to wrap the tailing and oplog interface in a pleasant API. rwynn was a large help with improving GTM's performance with varying levels of consistency guarantees.