Frontdesk

Services

The decision worth arguing about

Inbound mail lands through an SES receipt rule that writes the raw MIME to S3 and notifies SNS, and the non-obvious decision is how a reply gets attached to the right ticket. The standard trick is plus-addressing — support+1234@example.com — and it breaks in the field more often than vendors admit: distribution lists rewrite the envelope, some corporate gateways strip subaddressing entirely, and a customer forwarding the thread to a colleague loses it altogether. So Frontdesk threads primarily on RFC 5322 Message-ID, In-Reply-To and References, which survives forwarding and rewriting because it lives in the headers rather than the address. The cost is a table of every Message-ID the system has ever emitted, indexed and never pruned, which for a desk handling 40,000 tickets a month is roughly two million rows a year — small, but it must be treated as permanent, because pruning it is how you get a reply to a two-year-old ticket opening a new one. Two fallbacks sit behind it: a signed token in the subject line, and finally a requester-plus-subject-similarity match that creates a linked ticket rather than guessing. SLA timers run as a single sweeper scanning an indexed breach_at column every 30 seconds rather than a scheduled event per ticket, which caps breach detection granularity at half a minute and costs nothing at any ticket volume.

Services

The decision worth arguing about

Azure gives you a better mechanism for SLA timing than AWS does, and it is worth using rather than porting the sweeper across unchanged. Service Bus supports scheduled messages with an arbitrary future enqueue time, so each ticket can carry its own one-shot timer that fires exactly at the breach moment — no polling, no 30-second granularity, no sweeper to fall behind during a queue backlog. SQS caps delivery delay at 15 minutes, which is why that shape is unavailable on AWS without either EventBridge Scheduler at one schedule per ticket or the sweeper we actually use. The trade-off is bookkeeping: cancelling or moving a scheduled message requires holding its sequence number, so every SLA pause, reassignment or policy change has to find and cancel the old message and enqueue a new one, and a dropped sequence number leaves an orphan timer that fires on a ticket already closed. Frontdesk stores the sequence number on the SLA clock row and treats a fired timer for a resolved ticket as a no-op, which makes the failure mode harmless rather than merely unlikely. We run this on Service Bus Standard; Premium's fixed monthly floor is only worth it past roughly 1,000 messages a second, which no support desk of this size reaches.