There is no difference between Time and any of the three dimensions of Space except that our consciousness moves along it.
-The Time Machine - H. G. Wells
Filby is like source control for reference data but at runtime.
- With SCM systems like git, you have a commit log comprising a sequence of commits. Each commit is uniquely identified by a commit hash and typically bundles related source code changes.
- With Filby, you have a change log comprising a sequence of change sets. Each change set is uniquely identified by a change set id and typically bundles related reference data changes.
Like checking out a commit, your applications can use Filby to retrieve reference data for a given change set id. They can also inspect the changelog to find which change set was in effect at a given point in time, and subscribe to reference data update notifications.
See the example applications for usage.
Most applications require reference data that changes slowly over time. This presents the following challenges in a distributed or microservice architecture.
| Challenge | Notes |
|---|---|
| Consistency | Whenever we duplicate our reference data, we increase the likelihood of inconsistency. Even if we have one authoritive source of truth, we may cache the reference data in multiple systems, resulting in temporary inconsisenty unless cache updates are sychronised. Given the reference data is slow moving, a short period of inconsistency may be acceptable. |
| Load Times | Some reference data sets may be too large to desirably load over a network connection for web and mobile applications. Therefore we should discourage accidentally including large data sets into a client bundle, or requesting large data sets over a network. |
| Reliability | Requesting data sets over a network may fail, especially when mobile. Bundling local copies of reference data into the application (providing they are not too large) will aleviate this, but increase the potential for stale data. |
| Stale Data | Even though reference data is slow moving, it will still change occasionally. Therefore we need a strategy for refreshing reference data. |
| Temporality | When reference data changes, the previous values may still be required for historic comparisons. Therefore all reference data should have an effective date. Effective dates can also be used to synchronise updates by including future records when the values are known in advance. This comes at the cost of increased size, and there may still be some inconsistency due to clock drift and cache expiry times. |
| Evolution | Both reference data, and our understanding of the application domain evolves over time. We will at some point need to make backwards incompatible changes to our reference data, and will need to do so without breaking client applications. This suggests a versioning and validation mechanism. The issue of temporality compounds the challenge of evolution, since we may need to retrospecively add data to historic records. In some cases this data will not be known. |
| Local Testing | Applications may be tested locally, and therefore any solution sould work well on a local development machine. |
Solving such a complex problem becomes simpler when broken down. This project provides a server side framework for managing temporal reference data. In the following diagram, the mechanism for defining, loading, accessing and receiving notifications about reference data are provided by this framework. The RESTful API and Webhook must be manually created by the application developer. Example applications are provided to demonstrate how.
Change
Notification Webhook
┌─────────┐ ┌──────────────────────────────────────────────────────────────────┐
│ │ │ │
│ ▼ │ ▼
┌────────┐ ┌───────────┬──────────┐ GET /api/changelog?projection=$p&version=$v ┌──────────┐
│ │ │ │ ├──────────────────────────────────────────────────────────│ │
│ │ │ │ Example │ │ │
│ DB │◀────▶│ Filby │ App │ │ Client │
│ │ │ │ │ GET /api/projection/:version/:projection?changeSetId=$c │ │
│ │ │ │ ├──────────────────────────────────────────────────────────│ │
└────────┘ └───────────┴──────────┘ └──────────┘
▲
│
│
┌─────────────────────────────────────────────┐
│ │
│ Reference Data │
│ Change Sets │
│ │
└─────────────────────────────────────────────┘
▲ ▲ ▲ ▲
│ │ │ │
│ │ │ │
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│ │ │ │ │ │ │ │
│ CSV │ │ YAML │ │ JSON │ │ SQL │
│ │ │ │ │ │ │ │
└─────────┘ └─────────┘ └─────────┘ └─────────┘
The first of the two API calls, namely /api/changelog discloses the changes undergone by a projection (a view of the reference data), and provides a set of ids for requesting the projection at a point in time.
GET /api/changelog?projection=park&version=1
[
{
"changeSetId": 1,
"effective": "2019-01-01T00:00:00.000Z",
"description": "Initial park data",
"lastModified": "2018-12-01T16:20:34.383Z",
"eTag": "142475a5eddeec8f0786"
},
{
"changeSetId": 2,
"effective": "2020-01-01T00:00:00.000Z",
"description": "Park Calendars - 2020",
"lastModified": "2029-12-01T14:49:34.405Z",
"eTag": "a3dc15aa8d59d26e349d"
}
]The second API call, namely /api/projection/:version/:projection will return the reference data, correct at the time of the given changeSetId.
GET /api/projection/v1/park?changeSetId=2
[
{
"code": "DC",
"name": "Devon Hills"
},
{
"code": "GA",
"name": "Greenacres"
},
{
"code": "PV",
"name": "Primrose Valley"
}
]The example applications include a convenience feature where if the changeSetId is omitted, the client will be redirected to projection at the current change set.
GET /api/projection/v1/park
HTTP/1.1 307 Temporary Redirect
location: /api/projection/v1/park?changeSetId=8
At first glance, accessing projections via change sets may seem an unnecessary overhead, however it provides a number of worthy benefits.
- Reference data can be safely released (and potentially pre-cached) ahead of time.
- Clients can receive consistent reference data by fixing the changeSetId at the start of a transaction.
- Clients can receive the latest reference data by updating to the latest changeSetId when one becomes available.
- The changeSetId makes for an excellent cache buster, enabling the projection responses to be cached indefinitely.
- Projections provide multiple views of the same reference data, and can therefore be tailored to the needs of each client.
- Versioned projections supports backwards incompatible changes
Refering back to the previous list of challenges:
- Consistency is solved by the changeSetId mechanic.
- Load Times can be reduced because the solution is designed to be highly cacheable. Furthermore, response sizes can abe reduced by tailoring projections to the needs of the clients. Finally, network requests can be avoided completely by maintaining a local copy of the projected reference data.
- Stale Data can be prevented through use of the effective date, deploying reference data ahead of time and checking the changelog for updates. Clients can also be advised of updates via change notifications
- Historic data still accessible by using a previous changeSetId.
- Evolution of reference data is supported by versioned entities and projections.
- Local Testing is possible through HTTP mocking libraries.
Filby has the following important concepts
┌─────────────────┐
│ │
│ │ announces changes via
│ Projection │───────────────────────────┐
│ │ │
│ │ │
└─────────────────┘ │
│ depends on │
│ │
│ │
│ │
│ │
╱│╲ ╱│╲
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ │ │ │ │ │
│ │ │ │╲ delivered via │ │
│ Entity │ │ Notification │───────────────────│ Hook │
│ │ │ │╱ │ │
│ │ │ │ │ │
└─────────────────┘ └─────────────────┘ └─────────────────┘
│ aggregates ╲│╱ is raised by
│ │
│ │
│ │
│ │
╱│╲ │
┌─────────────────┐ ┌─────────────────┐
│ │ │ │
│ │╲ is grouped by │ │
│ Data Frame │──────────────────│ Change Set │
│ │╱ │ │
│ │ │ │
└─────────────────┘ └─────────────────┘
A Projection is a versioned view of one or more Entities, made available in the example application via a RESTful API. The implementor is responsible for writing the the projections, which will mostly be simple database to JSON transformations.
An Entity represents reference data. Entities may be stand alone, or form an object graph. We use a holiday park as an example, in which a 'Park' entity has many 'Season' entities. If you are familiar with event sourcing, they are implemented as an aggregate of one or more Data Frames. The dependency between projections and entities must be explicitly stated so we can emit Notifications when a new data frame is added.
A Data Frame is a snapshot of an entity, associated with a Change Set. There are two types of data frame, 'POST' which adds a new snapshot at a point in time, and 'DELETE' which indicates the entity has been deleted.
A Change Set groups a set of data frames (potentially for different entities) into a single unit with a common effective date.
Notifications are published whenever a new data frame is created. By subscribing to the notifications that are emitted per projection when the backing data changes, downstream systems can maintain copies of the data, with reduced risk of it becoming stale. For example, the client in the above diagram could be another backend system, caching proxy, a web application, a websocket application, a CI / CD pipeline responsible for building a client side data module, or an ETL process for exporting the reference data to the company data lake.
Notifications are retried a configurable number of times using an exponential backoff algorithm. It is safe for multiple instances of the framework to poll for notifications concurrently.
A Hook is an event the framework will emit whenever a projection is added, dropped or updated via a change set. Your application can handle these events how it chooses, e.g. by making an HTTP request, or publishing a message to an SNS topic. Unlike node events, the handlers are asynchronous. It is advised not to share hooks between handlers since if one handler fails but another succeeds the built in retry mechanism will re-notify both handlers.
All of above objects (Projections, Entities, Data Frames, etc) are defined using a domain specific language, which is dynamically converted into SQL and applied using a database migration tool called Marv. Whenever you need to make an update, simply create a new JSON or YAML migration file in the configurable migrations folder. A JSON schema is available in /lib/schema.json
# migrations/0001.define-park-schema.yaml
# Add enums for your reference data
# Equivalent of PostgreSQL's CREATE TYPE statement
- operation: ADD_ENUM
name: season_type
values:
- Owners
- Guests
# Adding entities performs the following:
# 1. Inserts a row into the 'fby_entity' table,
# 2. Creates a table 'park_v1' for holding reference data
# 3. Creates an aggregate function 'get_park_v1_aggregate' to be used by projections
- operation: ADD_ENTITY
name: park
version: 1
fields:
- name: code
type: TEXT # Any valid PostgreSQL type (including enums) are supported
- name: name
type: TEXT
identified_by:
- code # Used to aggregate the data frames
# Creates PostgreSQL check constraints
# This must be explicitly enabled in config since they introduce the potential for SQL Injection
checks:
park_code_len: LENGTH(code) >= 2
# Defining projections and their dependent entities
# Filby uses the dependencies to work out what projections are affected by reference data updates
- operation: ADD_PROJECTION
name: parks
version: 1
dependencies:
- name: park
version: 1
- name: season
version: 1
# A hook defines an asynchronous event that will be emitted by the framework whenever the
# reference data the projection changes.
# This is a projection specific hook which fires when the data
# supporting the specified projection changes
- operation: ADD_HOOK
name: sns/add-change-set/parks-v1 # Must be unique
event: ADD_CHANGE_SET
projection: Parks
version: 1
# This is a general hook that fires when the data
# supporting any projection changes
- operation: ADD_HOOK
name: data-lake/add-change-set/* # Must be unique
event: ADD_CHANGE_SET# migrations/0002.initial-data-load.yaml
# Add a change set containing one or more data frames for the previously defined entities
- operation: ADD_CHANGE_SET
description: Initial Data
effective: 2019-01-01T00:00:00Z
frames:
- entity: park
version: 1
action: POST
data:
# Adds a data frame for Devon Cliffs
- code: DC
name: Devon Cliffs
# Adds a data frame for Primrose Valley
- code: PV
name: Primrose Valley
- entity: park
version: 1
# Adds a data frame that will delete the entity identified
# by code HF from the effective data
action: DELETE
data:
- code: HF
# YAML can get verbose, so you can also import data frames from a local CSV
# The CSV requires a header row, starting with the action column and
# followed by the column names of your entity. When deleting data you
# only need to include the fields that identify the entity
- entity: park
version: 1
source: ./data/park-data-2019.csv
# action,code,name
# POST,KC,Kent Coast
# POST,CA,Caistor
# DELETE,TP,# migrations/0003.drop-unused-objects.yaml
# Deletes the specified enum
# Fails if the enums are still use
- operation: DROP_ENUM
name: season_type
# Deletes the specified entity
# Fails if the entities are depended on by projections or used in change sets
- operation: DROP_ENTITY
name: park
version: 1
# Deletes the specified projection and associated hooks
# Fails if the entities are still depended on by projections
- operation: DROP_PROJECTION
name: parks
version: 1
# Deletes the specified hook and associated notifications
- operation: DROP_HOOK
name: sns/add-change-set/park-v1The DSL deliberately omits the ability to drop change sets and data frames since they are supposed to be immutable. However, since Marv works with SQL, you can also use it for regular database migrations.
-- migrations/0004.delete-all-the-things.sql
START TRANSACTION;
-- deletes frame sets by entity
DELETE FROM fby_data_frame f
INNER JOIN fby_entity e ON e.id = f.entity_id;
WHERE e.name = 'Park'
AND e.version = 1;
-- deletes a change set by id
DELETE FROM fby_change_set
WHERE id = 4;
END TRANSACTION;In addition to deleting change sets, this can be useful for adding custom views or functions over the aggregated data frames to make your projections more efficient. Simply create a SQL migration file in the migrations directory instead of a YAML or JSON one.
-- migrations/0005.create-get-park-v1-function.sql
CREATE FUNCTION get_park_v1(p_change_set_id INTEGER)
RETURNS TABLE (code TEXT, name TEXT, season_type season_type, season_start TIMESTAMPTZ, season_end TIMESTAMPTZ)
AS $$
BEGIN
RETURN QUERY
SELECT p.code, p.name, s.type AS season_type, s.start AS season_start, s.end AS season_end
FROM get_park_v1_aggregate(p_change_set_id) p
LEFT JOIN get_season_v1_aggregate(p_change_set_id) s ON s.park_code = p.code
ORDER BY p.code ASC, s.start DESC, s.type ASC;
END;
$$ LANGUAGE plpgsql IMMUTABLE;PRO TIP: Since the results for a given change set should never change, consider making PostgreSQL functions IMMUTABLE so the output can be cached.
Filby provides a set of lifecycle methods and an API for retrieving change sets and projections, and for executing database queries.
Constructs a new Filby instance
Connects to the database and runs schema/data migrations
Stops polling for notifications then disconnects from the database
Starts polling the database for notifications
Stops polling the database for notifications, and waits for any inflight notifications to complete.
Under the hood, Filby uses eventemitter2 which unlike node's EventEmitter, supports asynchronous events. You can use these to listen for hooks and perform of asynchronous tasks like making an HTTP request for a webhook. The hook name is user defined and must be specified in the Hook Data Definition. The sole callback parameter is the Notification context (see the TypeScript definitions), e.g.
filby.subscribe('VAT Rate changed', async (notification) => {
await axios.post('https://httpbin.org/status/200', notification);
});If your event handler throws an exception it will be caught by Filby and the notifiation retried up to a configurable number of times, with an incremental backoff delay. If the maximum attempts are exceeded then Filby emits dedicated hook, Filby.HOOK_MAX_ATTEMPTS_EXHAUSTED. The sole callback parameter is the ErrorNotification context (see the TypeScript definitions), e.g.
filby.subscribe(Filby.HOOK_MAX_ATTEMPTS_EXHAUSTED, (notification) => {
console.error('Hook Failed', notification);
console.error('Hook Failed', notification.err.stack); // Careful not to log secrets that may be on the error
});Unsubscribes the specified handler, or all handlers if none is specified, from the specified event
Unsubscribes all handlers from all events
Returns the list of projections.
Returns the specified projection.
Returns the change log (an ordered list of change sets) for the given projection.
Returns the current changeset for the given projection based on the database's current time.
Returns the specified change set
Returns aggreated entity data for the specified changeSetId. The sort order will be in order of the entity's identifier fields (ascending, nulls last).
async function getParks(changeSetId) {
const rows = await filby.getAggregates(changeSetId, 'Park', 1);
return rows.map(toParkStructure);
};Passes a transactional node-pg client to the given callback. Use this to directly query the aggregate entity data for your projections. The aggregates are accessible via functions with the signature get_${entity}_v{version}_aggregate(changeSetId INTEGER). Entity names will be converted to lowercase and spaces converted to underscores. E.g.
async function getParks(changeSetId) {
return filby.withTransaction(async (client) => {
const { rows } = await client.query(`
SELECT p.code, p.name, s.type AS season_type, s.start AS season_start, s.end AS season_end
FROM get_park_v1_aggregate($1) p
LEFT JOIN get_season_v1_aggregate($1) s ON s.park_code = p.code
ORDER BY p.code ASC, s.start DESC, s.type desc ASC;
`, [changeSetId]);
return rows.map(toParkStructure);
});
};{
// All the database configuration is passed through to https://www.npmjs.com/package/pg
"database": {
"user": "fby_example",
"database": "fby_example",
"password": "fby_example"
},
// Specifies the path to the migrations folders and their permitted operations.
// Use this to restrict operations for non-techincal/less trusted users
// Defaults to [{ path: "migrations", permissions: ["ALL"] }].
// See the directory permissions table below for more details.
"migrations": [
{
path: "path/to/raw/migrations",
permissions: ["SQL"]
},
{
path: "path/to/schema/migrations",
permissions: ["ALL_OPERATIONS", "CHECK_CONSTRAINT"]
},
{
path: "path/to/data/migrations",
permissions: ["ADD_CHANGE_SET"]
}
],
"migrations": {
// Specifies the path to the migrations folder. Defaults to "migrations"
"directory": "path/to/migrations/folder"
},
"notifications": {
// The frequency Filby will check for new notifications
// Defaults to 1 minute
"interval": "5s",
// The initial delay before Filby starts checking for notifications
// (you still have to call filby.startNotifications)
// Defaults to 10s
"intialDelay": "1s",
// THe maximum number of times Filby will attempt to deliver a hook
// Defaults to 10
"maxAttempts": 20,
// The maximum amount of time Filby will wait before retrying a failed hook
// Defaults to 1 hour
"maxRescheduleDelay": "30s"
}
}| Permission | Notes |
|---|---|
| ADD_ENUM | Permits the ADD_ENUM operation. |
| ADD_ENTITY | Permits the ADD_ENTITY operation, but without check constraints. |
| ADD_HOOK | Permits the ADD_HOOK operation. |
| ADD_PROJECTION | Permits the ADD_PROJECTION operation. |
| ADD_CHANGE_SET | Permits the ADD_CHANGE_SET operation. |
| DROP_ENUM | Permits the DROP_ENUM operation. |
| DROP_ENTITY | Permits the DROP_ENTITY operation. |
| DROP_PROJECTION | Permits the DROP_PROJECTION operation. Deletions still cascade to hooks even if the DROP_HOOK permission is not granted. |
| DROP_HOOK | Permits the DROP_HOOK operation. |
| CHECK_CONSTRAINT | Permits the use of check constraints. Check contraints cannot be escaped or validated, and are therefore a potential source of SQL injection. |
| ALL_OPERATIONS | Permits all operations, however does permit check constraints or raw SQL migrations. |
| ALL | Permits all operations, check constraints and raw SQL migrations. |
This repository includes JavaScript and TypeScript example applications based on a Caravan Park business.
git clone [email protected]:acuminous/filby.git
cd filby
npm i
npm run dockercd examples/typescript
npm i
npm startcd examples/javascript
npm i
npm startOnce successfully started you should see the following output.
Server is listening on port 3000
See http://localhost:3000/documentation
Use CTRL+D or kill -TERM 18964 to stop
The applications include swagger documentation for the APIs but for a headstart try the following
GET http://localhost:3000/api/changelog?projection=park&version=1
This will return the changelog for the applications 'Park' projection.
[
{
"id":1,
"effective":"2019-01-01T00:00:00.000Z",
"description":"Initial Data",
"lastModified":"2024-01-08T07:37:55.415Z",
"entityTag":"e008f04be41843dd58cb"
},
{
"id":2,
"effective":"2020-01-01T00:00:00.000Z",
"description":"Park Calendars - 2020",
"lastModified":"2024-01-08T07:37:55.434Z",
"entityTag":"fc32163580c134bec0c0"
},
{
"id":3,
"effective":"2021-01-01T00:00:00.000Z",
"description":"Park Calendars - 2021",
"lastModified":"2024-01-08T07:37:55.444Z",
"entityTag":"c4240f6bab231fd059e2"
},
{
"id":4,
"effective":"2021-04-01T00:00:00.000Z",
"description":"Add Richmond",
"lastModified":"2024-01-08T07:37:55.458Z",
"entityTag":"a40672de82626c2518ef"
},
{
"id":5,
"effective":"2021-06-01T00:00:00.000Z",
"description":"Rename Richmond to Skegness",
"lastModified":"2024-01-08T07:37:55.466Z",
"entityTag":"d9766661224782a96ca6"
},
{
"id":6,
"effective":"2022-01-01T00:00:00.000Z",
"description":"Park Calendars - 2022",
"lastModified":"2024-01-08T07:37:55.476Z",
"entityTag":"841b9acde0c757decc20"
},
{
"id":7,
"effective":"2022-05-01T00:00:00.000Z",
"description":"Delete Greenacres",
"lastModified":"2024-01-08T07:37:55.489Z",
"entityTag":"56aae19316c7eed13adb"
},
{
"id":8,
"effective":"2023-01-01T00:00:00.000Z",
"description":"Park Calendars - 2023",
"lastModified":"2024-01-08T07:37:55.495Z",
"entityTag":"25c02cea536038f310e7"
}
]Now pick a change set id and request the Park data at that point in time...
GET http://localhost:3000/api/projection/v1/park?changeSetId=1
[
{
"code":"DC",
"name":"Devon Cliffs",
"seasons":[
{
"type":"Owners",
"start":"2019-03-01T00:00:00.000Z",
"end":"2019-11-30T00:00:00.000Z"
},
{
"type":"Guests",
"start":"2019-03-15T00:00:00.000Z",
"end":"2019-11-15T00:00:00.000Z"
}
]
},
{
"code":"GA",
"name":"Greenacres",
"seasons":[
{
"type":"Owners",
"start":"2019-03-01T00:00:00.000Z",
"end":"2019-11-30T00:00:00.000Z"
},
{
"type":"Guests",
"start":"2019-03-15T00:00:00.000Z",
"end":"2019-11-15T00:00:00.000Z"
}
]
},
{
"code":"PV",
"name":"Primrose Valley",
"seasons":[
{
"type":"Owners",
"start":"2019-03-01T00:00:00.000Z",
"end":"2019-11-30T00:00:00.000Z"
},
{
"type":"Guests",
"start":"2019-03-15T00:00:00.000Z",
"end":"2019-11-15T00:00:00.000Z"
}
]
}
]