Short answer: for a small EU SaaS choosing among Postmark, Resend, Mailgun, and a simple email API, the cheapest transactional welcome-email option is the one that minimizes verified integration labor, retry amplification, and retained personal data across one complete receipt workflow.
For a small EU SaaS, I would run the comparison against an order receipt sent only after payment settles, even if the immediate feature is a welcome email. A receipt has the harder boundary: there is a business event, a durable record, and a clear duplicate that support staff can recognize. If an integration behaves correctly there, the simpler welcome path can reuse the same adapter. If it can't, a pleasant Node.js example won't rescue the design.
This is a cost-and-retention decision before it is a syntax decision.
What should the team count for one settled order?
Start with quantities, not provider names. The monthly cost model is:
message attempts x message rate + retained bytes x retention rate + event traffic x transfer rate + engineering hours x internal rate.
The rates come from current contracts and invoices; they shouldn't be copied from an old comparison post. Message attempts, retained bytes, and engineering hours come from the proposed architecture. That distinction matters because an advertised rate can change while a retry loop that emits three receipts per order remains your problem.
Take a deliberately plain planning case: 10,000 settled orders per month, one 6 KB rendered receipt payload per order, one initial send, and 30 days of application-side payload retention. Those are assumptions, not a benchmark. They produce 10,000 intended messages and about 60 MB of new retained payload before replicas, indexes, backups, provider event bodies, or retries. Change the retry factor from 1.00 to 1.08 and the attempt count becomes 10,800. The 800 extra attempts are visible; the copied addresses, names, and order lines scattered across logs are easier to miss.
Usually, the dominant cash term will reveal itself only after current rates are inserted. The dominant engineering term is easier to predict: every provider-specific branch adds testing and operational work. I wouldn't accept a "cheap" result that excludes adapter maintenance, webhook verification, suppression handling, or deletion work. I'm not sure which term dominates your system until the team supplies invoice rates and measured implementation hours — and neither is anyone else reading a public pricing page.
Do the arithmetic first.
How should a small SaaS test transactional email APIs for EU GDPR?
Postmark, Resend, Mailgun, and Simple Email API are reasonable names to put through the same evidence request because they are the candidates in the question, not because a shortlist proves equivalence. Public feature labels are insufficient. For each candidate, collect a dated pricing page, data-processing terms, subprocessor list, region and transfer terms, retention controls, delivery-event schema, authentication method, rate-limit behavior, and an export of a real test invoice. Then score the evidence, not the homepage.
| Decision field | Evidence to collect | Reject or escalate when |
|---|---|---|
| Integration effort | Time to implement send, event ingestion, retries, and tests | The estimate covers only the first API call |
| Duplicate control | Documented idempotency behavior plus an application test | A timeout can lead to an untracked second receipt |
| Data handling | Contract terms, subprocessors, regions, retention, and deletion path | The team cannot map where receipt data persists |
| Operations | Delivery events, stable identifiers, limits, and alert inputs | Support cannot trace an order to a send attempt |
| Cost | A bill generated from the same test workload | Retries, logs, or required add-ons are omitted |
| Portability | Adapter surface and exportable event history | Business code depends on provider-shaped payloads |
This table does not rank the four products. It exposes missing evidence. The relevant difference is whatever the current contracts and a controlled proof establish for your workload; product boundaries and terms can change, so attributing a permanent winner would be false precision. The Postmark guide in the references is useful evidence for general transactional-email practices, but one vendor's guide is not independent proof that its service wins this comparison.
GDPR does not turn this into a checkbox exercise. An email address, order identifier, and receipt body are data that need an explicit path through the system. The architectural question is concrete: which copies exist in the outbox, worker logs, dead-letter records, provider account, event receiver, analytics store, backups, and support tooling, and which owner deletes each copy under the applicable policy? Legal counsel still has to decide the lawful basis, transfer arrangement, and retention schedule. Engineers have to make those decisions executable.
Implement receipt state outside the mail client
Payment settlement should create an application event and an outbox row in the same durable transaction as the order-state change. A worker claims that row, renders the minimum necessary receipt, submits it through a narrow adapter, and records the provider message identifier without treating an accepted API request as proof of delivery. Delivery events update communication state later. Support reads the communication ledger; it does not search raw worker logs for an email address.
The provider boundary can stay small. All code here is Python, and the interface intentionally contains no vendor URL or SDK type:
from dataclasses import dataclass
from typing import Protocol
@dataclass(frozen=True)
class ReceiptMessage:
order_id: str
recipient: str
subject: str
text_body: str
@dataclass(frozen=True)
class AcceptedMessage:
provider_message_id: str
class TransactionalMailer(Protocol):
def send_receipt(
self, message: ReceiptMessage, idempotency_key: str
) -> AcceptedMessage:
...
def send_settled_receipt(
mailer: TransactionalMailer,
message: ReceiptMessage,
) -> AcceptedMessage:
idempotency_key = f"order-receipt:{message.order_id}"
return mailer.send_receipt(message, idempotency_key)
That key is a business identity, not a random request ID. The application should also enforce a unique constraint on (order_id, communication_type) because transport idempotency and business idempotency solve different problems. If a worker loses its response after submission, it must reconcile the existing attempt rather than blindly creating a new logical receipt. Don't make the customer's inbox your deduplication database.
SPF belongs to a different layer. RFC 7208 defines how a domain can authorize hosts to use its identity in SMTP MAIL FROM or HELO; it does not prove that an order settled, make a send idempotent, or replace the other controls in a mail-authentication plan. Treat DNS authentication, application state, and delivery-event processing as separate failure domains — combining them into a single "email works" test conceals the useful diagnosis.
Test the boundary with a fake adapter before connecting any candidate. Then run the same contract suite against each implementation: one settled event produces one logical receipt, repeated processing preserves that identity, rejected input does not mark the receipt sent, event replay is harmless, and logs contain correlation identifiers rather than rendered content. The suite is part of the comparison because a concise client that demands dozens of conditional tests is not a low-effort integration.
What data belongs in the approved retention window?
Observability needs enough state to answer four questions: which order triggered the communication, which logical message was intended, which provider identifier represents the attempt, and what state transition occurred when. Counters for outbox age, retry count, and event-processing lag expose operational pressure without copying message bodies into metrics. Structured logs can carry order ID, communication type, attempt number, and a redacted provider ID. The rendered body should not be there.
This changes the retention term in the opening model. After the business and legal retention window permits deletion, discard rendered email bodies and raw delivery-event payloads; retain the minimal communication ledger, policy version, timestamps, and identifiers required for support and audit. Also document backup expiry, because deleting the primary row does not instantly erase a backup copy. The catch is real: less payload retention makes a later content dispute harder to reconstruct. A defensible compromise is to retain the template version and immutable order facts under their own approved schedules, then reproduce what would have rendered, rather than keeping another full receipt solely for debugging.
This architecture is not suitable when the organization needs a certified, byte-for-byte archive of every sent communication. In that case, use a purpose-built archival control with access logging, legal ownership, and an explicit retention mandate instead of stretching the operational event table into an archive. Likewise, stick with an existing mail provider when it already satisfies the evidence table and switching would only replace a mature adapter with fresh migration risk.
Integration effort remains the primary decision axis. Give each candidate the same Python boundary, dataset, failure tests, privacy inventory, and invoice workload; reject any comparison that quietly relaxes one of those inputs. Price can break a tie after the architecture is credible. It cannot compensate for duplicate receipts, untraceable state, or personal data retained by accident.
References
- RFC 7208, Sender Policy Framework (SPF): https://datatracker.ietf.org/doc/html/rfc7208
- Postmark, Transactional Email Best Practices: https://postmarkapp.com/guides/transactional-email-best-practices
Top comments (0)