Nerium

Nerium is a simple, lightweight, headless Flask-based query broker microservice that submits flexible SQL to any SQLAlchemy-supported database, and returns results as JSON or CSV—enabling rapid development and deployment of API endpoints for business intelligence reporting and data analytics applications.

Inspired by jamstack workflows, Nerium's queries and serialization formats are simply text files stored on the filesystem. Developers and report analysts can write SQL queries in their preferred editor, manage them with standard version control tools, and upload or mount them into a Nerium installation. In keeping with SQLAlchemy usage, query parameters can be specified in key=value format, and (safely!) injected into your query in :key format. Queries can also include Jinja template syntax, allowing greater flexibility and reuse of SQL logic.

Default JSON output represents data as an array of objects, one per result row, with database column names as keys. The default schema also provides top-level nodes for name, metadata, and params (details below). A compact JSON output format may also be requested, with separate column (array of column names) and data (array of row value arrays) nodes for compactness. Additional formats can be added by adding marshmallow schema definitions to format_files.

Nerium supports any backend database that SQLAlchemy can, but since none of these are hard dependencies, drivers aren't included in the default setup, and the Dockerfile only supports PostgreSQL. If you want Nerium to work with other databases, you can install Python connectors with uv add, either in a local project or by creating your own Dockerfile using FROM tymxqo/nerium. (To ease installation, options for nerium[mysql], nerium[pg], and nerium[snowflake] are provided as extras in pyproject.toml)

Install/Run

Using Docker

docker run -d --name=nerium \
--envfile=.env \
-v /local/path/to/query_files:/app/query_files
-v /local/path/to/format_files:/app/format_files \
-p 5000:5000 tymxqo/nerium

curl http://localhost:5000/v1/<query_name>?<params>

You might also want to use tymxqo/nerium as a base image for your own custom container, in order to add different database drivers, etc. Or you can build locally from the included Dockerfile. The base image includes the psycopg (v3) PostgreSQL adapter, along with gunicorn WSGI server for a production-ready service.

When building, pass the commit being built so the running instance can report it:

docker build --build-arg GIT_COMMIT=$(git rev-parse HEAD) -t nerium .

The / health endpoint returns that value as commit, alongside status and version, which makes it possible to tell exactly what code a deployed instance is running. Omit the build argument and commit reads unknown.

Local install

python -m pip install nerium[pg]

Or install latest source from Github:

git clone https://github.com/nerium-data/nerium.git
cd nerium
uv sync

(Nerium uses uv for dependency and environment management. uv sync creates a .venv from the pinned versions in uv.lock; prefix commands with uv run to use it. To include database drivers or the production server, add extras: uv sync --extra pg --extra serve.)

Then add a query_files (and, optionally, format_files) directory to your project, write your queries, and configure the app as described in the next section. The command FLASK_APP=nerium/app.py flask run starts a local development server running the app, listening on port 5000. For production use, you will want to add a proper WSGI server (we like gunicorn).

Configuration

DATABASE_URL for query connections must be set in the environment (or in a local .env file). This is the simplest configuration option.

For PostgreSQL, Nerium uses psycopg (v3), which SQLAlchemy does not select by default. Name the driver explicitly in the URL scheme — postgresql+psycopg://user@host:port/database — otherwise SQLAlchemy looks for the older psycopg2 and the connection fails.

Script file paths

By default, Nerium looks for query and format schema files in query_files and format_files respectively, in the current working directory from which the service is launched. QUERY_PATH and FORMAT_PATH environment variables can optionally be set in order to use files from other locations on the filesystem, as desired.

Snowflake keypair authentication

Snowflake connections can authenticate with an RSA keypair instead of a password. Install the Snowflake driver (nerium[snowflake], or uv sync --extra snowflake), set DATABASE_URL to a snowflake:// URL with no password, and supply the private key in the environment as either:

• PRIVATE_KEY_PATH — path to a PEM-formatted private key file, or
• PRIVATE_KEY — the PEM key itself. Because many config systems can't hold multi-line values, literal \n sequences in this variable are treated as newlines, so the key can be provided on a single line.

Set one or the other; if both are set, PRIVATE_KEY_PATH takes precedence. If the key is encrypted, also set PRIVATE_KEY_PASSPHRASE. The key is only read for snowflake:// URLs, and is passed to the driver by sqla-raw.

Multiple Data Sources

If you want to query multiple databases from a single Nerium installation, any individual query file can define its own database_url as a key in YAML front matter (see below). This will override the $DATABASE_URL setting in the environment for that query only. If you have a large number of queries across several databases, keep in mind that running a separate Nerium instance for each database could be an option.

Security and API key authorization

As a simple minimal microservice by design, Nerium's default security posture is to presume its clients and the network between them can be trusted, and that it is the responsibility of client applications and/or a separate authentication service to ensure that users can only request reports they are authorized to see.

If API_KEY is set in the environment, Nerium will check all incoming requests for an X-API-Key header and compare it to the envvar value. Requests missing the header receive a 400 error, and requests with an invalid value in the header get a 403. This can be used simply to afford an extra measure of internal security, or perhaps, for example, to run separate instances of Nerium in a multi-tenant configuration: each tenant receives its own key which is then matched with the setting in a particular Nerium deployment. These key values are to be protected as shared secrets, and shared with client operators via secure channels. We recommend using secrets.token_hex() or similar to generate them.

Usage

Query files

As indicated above, queries are simply text files placed in a local query_files directory, or another arbitrary filesystem location specified by QUERY_PATH in the environment. The base name of the file (stem in Python pathlib parlance) will determine the {query_name} portion of the matching API endpoint.

Query parameters

Use :<param> to specify bind parameters in your query text. Clients can then specify values for these bind parameters in their request, passed either as JSON or query string arguments.

Metadata

Query files can optionally include a YAML metadata frontmatter block. The use of a special comment for metadata allows for the SQL file to be used as-is in other SQL clients. To add this metadata, create a multiline comment surrounded by \* ... */ markers, and within this comment, surround the YAML document with standard triple-dashed lines, as in this example:

/*
---
Author: Joelle van Dyne
Description: Returns all active usernames in the system
---
*/
select username from user;

Metadata can generally be thought of as a way to pass arbitrary key-value pairs to a front-end client; in the default format, the metadata is simply returned as an attribute of the results response. (The compact formatter drops the metadata.) Other possible use cases include whatever a reporting service and front-end developer want to coordinate on.

There are a couple of special-case metadata items:

1. As noted above, it can be used to specify a database connection for the query, overriding the main DATABASE_URL in the environment
2. If the metadata includes a params block, its contents are returned as the params object in the v1/<query_name/docs discovery response.
3. Similarly, metadata describing columns will populate that section of the /docs/ response.

In the absence of explicit metadata, Nerium attempts to find column specifications and named parameters for /docs endpoints by inspecting the query text itself with sqlparse. Although it is more manual, a metadata comment can provide greater detail in these sections — a report developer might specify the data type of a column or parameter, for example, in addition to its name.

Jinja templating

Nerium supports Jinja templating syntax in queries. The most typical use case would be for adding optional filter clauses to a query so that the same SELECT statement can be reused without having to be repeated in multiple files, for example:

select username
     , user_id
     , display_name
  from user
{% if group %}
 where user.group = :group
{% endif %}

Jinja filters and other logic can be applied to inputs, as well.

WARNING!

The Jinja template is rendered in a SandboxedEnvironment, which should protect against server-side template injection and most SQL injection tactics. It should not be considered perfectly safe, however. Use this feature sparingly; stick with SQLAlchemy-style :key named parameters for bind value substitutions, and test your specific queries carefully. It should almost go without saying that database permission grants to the user Nerium connects as should be well-restricted, whether one is using Jinja syntax or not.

One known dangerous case is if your entire query file just does a Jinja variable expansion and nothing else: {{ my_whole_query }}. This will allow execution of arbitrary SQL and you should never make a template like this available.

Custom format files

For serialization formats besides the built-in default and compact, schema definitions can be added to your format_files directory, using the Python marshmallow library. Similarly to query files, the app will look for a format module name matching the {format} specified in the endpoint URL. The app expects a marshmallow.Schema subclass named ResultSchema. Available attributes passed to this schema are all found in the original query object. (See nerium/schema for examples of how this is done by built-in formats.)

API

Results endpoints

URLs

• /v1/<string:query_name>?<query_params>
• /v1/<string:query_name>/<string:format>?<query_params>

As shown above, query_name and format may be accessed as part of the URL structure, or can be passed as parameters to the request.

Because we're retrieving report results here, the request is a GET in any case, but parameters may be sent in a JSON body or as querystring parameters. [Yes, sending JSON with a GET feels weird to us too, but as we're fetching serialized results and not changing any data here, GET seems to be the most appropriate REST semantic. It does work, we promise!] Note that query_name and format from URL base path will be preferred, even if a request to such a path happens to include either key in the request body (client apps should avoid doing this to minimize confusion).

query_name should match the name of a given query script file, minus the file extension. URL querystring parameters (or JSON keys other than query_string and format) are passed to the invoked data source query, matched to any parameter keys specified in the query file. If any parameters expected by the query are missing, an error will be returned. Extra/unrecognized parameters are silently ignored (this might seem surprising, but it's standard SQLAlchemy behavior for parameter substitution).

format path may be included as an optional formatter name, and defaults to 'default'. Other supported formatter options are described in Content section below.

Unknown values of format will silently fall back to default.

Method

GET

Success Response

Code: 200

Content:

'default': {"name": "<query_name>", "data": [{<column_name>:<row_value>, etc..., }, {etc...}, ], "metadata": {<key>: <value>, etc..., }, "params": {<array of name-value pairs submitted to query with request>}}
'compact': {"columns": [<list of column names>], "data": [<array of row value arrays>]} 'csv': <csv formatted string (with \r\n newline)> 'csv.gz': <gzip compressed csv stream>

Of course, it is possible that a database query might return no results. In this case, Nerium will respond with an empty JSON array [] for the data attribute, regardless of specified format. This is not considered an error, and clients should be prepared to handle it appropriately.

Error Responses

Code: 400

Content: {"error": <exception.repr from Python>}

Documentation endpoints

Report listing endpoint

URLs

• /v1/docs/

Method

GET

Success Response

Code: 200

Content: {"reports": [<an array of report names>]}

Report description endpoints

URLs

• /v1/{string:query_name}/docs/

Method

GET

Success Response

{"columns":[<list of columns from report>],"error":false,"metadata":{<report: metadata object>},"name":"<query_name>","params":[<array of parameterized keys in query>],"type":"sql"}

Tests

pytest tests are located in tests/. uv sync installs test prerequisites along with everything else; then they can be run with: uv run coverage run -m pytest (or just uv run pytest).