Phased Deployment Guide

tobira is designed for zero-downtime integration with your existing mail infrastructure. This guide describes a phased rollout strategy.

Overview

Phase A: MTA standard filters (immediate)
    ↓
Phase B: ML model training (parallel)
    ↓
Phase C: ML model deployment (zero-downtime switch)
    ↓
Phase D: Continuous improvement

Phase A: MTA Standard Filters

Start with your MTA's built-in filtering (Bayesian, SPF, DKIM, fuzzy hashing, etc.). Install the tobira plugin but keep it disabled.

Goal: Begin collecting labeled spam/ham data for ML training.

# Install tobira and configure for your MTA
tobira init

tobira init auto-detects your running MTA and generates the appropriate plugin configuration.

Duration: Immediate. Start collecting data from day one.

Phase B: ML Model Training

While Phase A runs in production, train your ML model on the collected data.

1. Anonymize training data (PII removal):

   # tobira handles PII anonymization automatically during training
   

2. Train and evaluate:

   tobira evaluate --data /path/to/labeled-data.jsonl --backend onnx
   

3. Check if the model meets quality thresholds:

   tobira monitor --check-threshold --target-f1 0.95
   

Duration: Hours to days, depending on data volume and model complexity.

Requirement: F1 score > 0.95 before proceeding to Phase C.

Phase C: ML Model Deployment

Deploy the trained model with zero downtime.

1. Start the tobira API server:

   tobira serve --config /etc/tobira/tobira.toml
   

2. Verify the server is healthy:

   tobira doctor
   

3. Enable the MTA plugin:

   rspamdSpamAssassinHarakaPostfix milter

   Edit /etc/rspamd/local.d/tobira.conf and reload:

   rspamadm configtest && systemctl reload rspamd
   

   Enable the plugin in local.cf and restart:

   systemctl restart spamassassin
   

   Add tobira to config/plugins and restart:

   systemctl restart haraka
   

   Start the milter daemon:

   tobira milter --config /etc/tobira/milter.conf
   

Duration: Configuration change only. No mail service interruption.

Phase D: Continuous Improvement

Monitor model performance and retrain as needed.

# One-shot analysis
tobira monitor

# Continuous monitoring (daemon mode)
tobira monitor --watch

tobira's monitoring detects:

• Concept drift — Score distribution shifts over time
• Threshold tuning — Optimal classification thresholds
• Backend upgrade suggestions — When to switch to a more accurate backend

Infrastructure Requirements

Phase	Infrastructure	Data
A	MTA filter (rspamd / SpamAssassin / Haraka)	None (start collecting)
B	GPU environment (for training)	500+ labeled emails
C	API server (CPU is sufficient)	Trained model
D	Same as Phase C	Continuously collected data