Bigbucket

Bigbucket is a serverless NoSQL database with a focus on scalability, availability and simplicity. It has a Bigtable-style data model with storage backed by a Cloud Storage Bucket.

It's serverless in the sense of the storage layer being fully managed and the API layer being stateless and horizontally scalable, which makes it ideal to run in serverless offerings like Google Cloud Run/Functions or AWS Lambda. No servers/disks to manage, no masters/slaves, no sharding; just create a bucket and point a binary to it.

The goal of the project is to offer a simple, easy to manage wide column database, with a lower barrier to entry compared to Bigtable and Cassandra, to folks who care more about scalability and availability rather than maximum performance.

Note: The project is currently in alpha and under development. I would not recommend it for production usage just yet, but if you do try it, feedback would be highly appreciated :)

Features:

• Bigtable-style data model (wide column / two-dimensional KV)
• Storage backed by a Cloud Storage Bucket (GCS available, S3 planned)
• Fully stateless frontend with a simple RESTful API
• Horizontally scalable. Need more throughput? Just add more replicas and raise Cloud Storage quotas if necessary
• (WIP) Flexible data schema with the option to enforce at API layer
• (WIP) Authentication and access policies per operation, per table/column
• Async delete of tables and columns? Just run another instance in cleaner/garbage-collection mode
• Row operations(read/set/delete) are parallelized (e.g. 1 row read ~= 1k row read)
• Row cells compressed with Zstandard
• Out of the box from Cloud Storage:
• Strongly consistent writes
• Highly available and replicated objects
• Pay per request and per GB stored /month
• Object versioning and lifecycle management
• Auditing
• Single region for lower latency or multi-region for more availability

Sections:

• Architecture and data model
• API
• Clients
• Running
• Locally
• Cloud Run
• Kubernetes
• Configuration
• Contributing
• TODO / Ideas

Architecture and data model

Here's an overview diagram of how Bigbucket would look like deployed and serving requests on GCP. In this scenario:

• Bigbucket API allows clients to interact with the wide column store through a RESTful API, like listing/counting/reading/writing/deleting rows, and listing/deleting columns and tables. It's deployed as an auto-scaling private Cloud Run service, with appropriate bucket permissions. Client requests will be load balanced between the containers and authenticated using their service account identity token (GCP docs).

• Bigbucket Cleaner removes tables and columns that have been marked for deletion. It's deployed as a single-container private Cloud Run service, with appropriate bucket permissions, triggered every hour by a Cloud Scheduler job.

To run this yourself, check out the Running in Cloud section.

Also here's a visual representation of the data model and the terminology behind it. Compared to Bigtable, the current model has 2 key differences:

1. There are no column families, just columns, mainly for the sake of simplicity and because object partitioning is fully managed by the bucket. Column prefix filtering, in adition to the current specific column filtering, can be added later if there's a need for it.

2. The initial version of the API only allows fetching of the latest cell version, but there's a plan to allow listing/reading previous cell values leveraging the object versioning feature of the bucket. This allows users to have full control over the max number of versions allowed or specify an expiry/delete lifecycle, using cloud provider tools already available and understood.

A few things to keep in mind when designing your schema are:

• Row keys are sorted and they are the only way to filter your rows. The value of cells cannot be queried, only returned. Design your row keys with your queries in mind, with the more important values first, taking into account key prefixes as that is currently the only way to scan the table. The Cloud Bigtable guide to choosing row keys is a good resource here.
• Reading a single row with a specified key and columns is the fastest way to get a cell value. Using row key prefixes or not specifying which columns you want will require additional requests to the bucket.
• The cells are compressed with Zstandard, so no need to pre-compress yourself.
• It's cheaper and faster to fetch 10 big columns rather than 100 small columns. Try to combine similarly read data together in the same column.
• Write-heavy data should be kept in separate columns as there is an update limit of once per second for the same cell (GCS quotas).

API

Note on naming: Tables, columns and row keys follow object name requirements from Google Cloud Storage. In short, Bigbucket API will return "HTTP 400 Bad Request" when trying to use tables, columns or row keys starting with dot "." or containing: \n, \r, \t, \b, #, [, ], *, ?, /

Table

Endpoint: /api/table

Create tables

Schema is flexible by default so tables are automatically created when a new row is inserted. Schema enforcement is planned for the near future.

List tables

curl -X GET "http://localhost:8080/api/table"

Response:
{
  "tables": [
    "test"
  ]
}

Delete table

Tables marked for deletion will need to be garbage-collected by running Bigbucket in cleaner mode. See Running section below.

Querystring parameters:

  table (required)

curl -X DELETE "http://localhost:8080/api/table?table=test"

Response:
{
  "success": "Table 'test' marked for deletion"
}

Column

Endpoint: /api/column

Create columns

Schema is flexible by default so columns are automatically created when rows are inserted. Schema enforcement is planned for the near future.

List columns

Because of schema flexibility, this will list the columns from only the first row in your table.

Querystring parameters:

  table (required)

curl -X GET "http://localhost:8080/api/column?table=test"

Response:
{
  "columns": [
    "col1",
    "col2",
    "col3"
  ],
  "table": "test"
}

Delete column

Columns marked for deletion will need to be garbage-collected by running Bigbucket in cleaner mode. See Running section below.

Querystring parameters:

  table  (required)
  column (required)

curl -X DELETE "http://localhost:8080/api/column?table=test&column=col1"

Response:
{
  "success": "Column 'col1' marked for deletion in table 'test'"
}

Row

Endpoint: /api/row

Count rows

Querystring parameters:

  table  (required)

  prefix (optional) // Row key prefix

curl -X GET "http://localhost:8080/api/row/count?table=test"

Response:
{
  "rowsCount": "5",
  "table": "test"
}

List row keys

Querystring parameters:

  table  (required)

  prefix (optional) // Row key prefix

curl -X GET "http://localhost:8080/api/row/list?table=test"

Response:
{
  "rowKeys": ["key1", "key2", "key3", "key4", "key5"],
  "table": "test"
}

Read rows

Querystring parameters:

  table   (required)

  columns (optional) // Comma separated
  limit   (optional) // Limit of rows returned

  Exclusive (only one of):

  key     (required) // Row key
  prefix  (required) // Row key prefix

Read single row with specified columns (fastest read op):

curl -X GET "http://localhost:8080/api/row?table=test&key=key1&columns=col1,col2,col3"

Response:
{
  "key1": {
    "col1": "val",
    "col2": "val",
    "col3": "val"
  }
}

Read all rows:

curl -X GET "http://localhost:8080/api/row?table=test"

Response:
{
  "key1": {
    "col1": "val",
    "col2": "val",
    "col3": "val"
  },
  "key2": {
    "col1": "val",
    "col2": "val",
    "col3": "val"
  },
  ...
}

Read rows with prefix and limit:

curl -X GET "http://localhost:8080/api/row?table=test&prefix=key&limit=1"

Response:
{
  "key1": {
    "col1": "val",
    "col2": "val",
    "col3": "val"
  }
}

Set row

Querystring parameters:

  table   (required)
  key     (required) // Row key

JSON Payload:

  {
    column (string): value (string),
  }

curl -X POST "http://localhost:8080/api/row?table=test&key=key5" \
  -d '{"col1": "newVal", "col3": "newVal"}'

Response:
{
  "success": "Set row key 'key5' in table 'test'"
}

Delete rows

Querystring parameters:

  table   (required)

  Exclusive (only one of):

  key     (required) // Row key
  prefix  (required) // Row key prefix

Delete one row:

curl -X DELETE "http://localhost:8080/api/row?table=test&key=key5"

Response:
{
  "success": "Row with key 'key5' was deleted from table 'test'"
}

Delete rows with prefix:

curl -X DELETE "http://localhost:8080/api/row?table=test&prefix=key"

Response:
{
  "success": "4 rows with key prefix 'key' were deleted from table 'test'"
}

Clients

• Python3
• Golang

Running

Create a GCS bucket

gsutil mb -p <your-project> -l EUROPE-WEST1 gs://<bucket-name>/

Running locally

Binary

git clone https://github.com/adrianchifor/Bigbucket
cd Bigbucket
make

API

./bin/bigbucket --bucket gs://<bucket-name>

Cleaner

./bin/bigbucket --bucket gs://<bucket-name> --cleaner --cleaner-interval 30

Docker

API

docker run -d --name "bigbucket-api" \
  -e BUCKET=gs://<bucket-name> \
  -v ${HOME}/.config/gcloud:/root/.config/gcloud \
  -p 8080:8080 \
  ghcr.io/adrianchifor/bigbucket:latest

Cleaner

docker run -d --name "bigbucket-cleaner" \
  -e BUCKET=gs://<bucket-name> \
  -e CLEANER=true \
  -e CLEANER_INTERVAL=30 \
  -v ${HOME}/.config/gcloud:/root/.config/gcloud \
  ghcr.io/adrianchifor/bigbucket:latest

We mount ${HOME}/.config/gcloud:/root/.config/gcloud in both cases so the containers can use our local gcloud credentials to talk to the bucket.

Let's test it

# Set a row
$ curl -X POST "http://localhost:8080/api/row?table=test&key=key1" \
  -d '{"foo": "hello", "bar": "world"}' | jq .

{
  "success": "Set row key 'key1' in table 'test'"
}

# Get the row
$ curl -X GET "http://localhost:8080/api/row?table=test&key=key1" | jq .

{
  "key1": {
    "bar": "world",
    "foo": "hello"
  }
}

# Delete the table
$ curl -X DELETE "http://localhost:8080/api/table?table=test" | jq .

{
  "success": "Table 'test' marked for deletion"
}

# Check cleaner for garbage-collection
$ docker logs bigbucket-cleaner

2020/05/04 19:10:34 Running cleaner...
2020/05/04 19:10:35 Running cleaner every 30 seconds...
2020/05/04 19:18:36 Table 'test' cleaned up

# Stop containers
$ docker kill bigbucket-api
$ docker kill bigbucket-cleaner

Running in Cloud

There's an example deploy/run.yaml file on how to deploy Bigbucket to GCP, like in the first diagram in Architecture and data model, using run-marathon.

Modify deploy/run.yaml to suit your environment

git clone https://github.com/adrianchifor/Bigbucket
cd Bigbucket/deploy
vim run.yaml

Replace all occurances of your_* with your own project, region, bucket etc.

Let's deploy it

# Install run-marathon
$ pip3 install --user run-marathon

$ run check
Cloud Run, Build, Container Registry, PubSub and Scheduler APIs are enabled. All good!

# Build and push Docker image to GCR
$ run build
...

# Create service accounts, attach IAM roles, deploy to Cloud Run and create Cloud Scheduler job
$ run deploy
...

# Get the endpoint of your API
$ run ls
   SERVICE        REGION        URL                        LAST DEPLOYED BY   LAST DEPLOYED AT
✔  bigbucket-api  europe-west1  https://YOUR_API_ENDPOINT  you                some time

Let's test it

# Use your account identity token to authenticate to the private API endpoint
$ alias gcurl='curl --header "Authorization: Bearer $(gcloud auth print-identity-token)"'

# Set a row
$ gcurl -X POST "https://YOUR_API_ENDPOINT/api/row?table=test&key=key1" \
  -d '{"foo": "hello", "bar": "world"}' | jq .

{
  "success": "Set row key 'key1' in table 'test'"
}

# Get the row
$ gcurl -X GET "https://YOUR_API_ENDPOINT/api/row?table=test&key=key1" | jq .

{
  "key1": {
    "bar": "world",
    "foo": "hello"
  }
}

Nice! Now you've got load balanced, auto-scaling private Bigbucket API containers with TLS, and a Bigbucket Cleaner container triggered every hour.

Running in Kubernetes

Change the BUCKET environment variable in deploy/k8s.yaml and run:

$ kubectl apply -f deploy/k8s.yaml

deployment.apps/bigbucket configured
service/bigbucket configured
cronjob.batch/bigbucket-cleaner configured

This will deploy the