Stage 5 of 12
Sending Infrastructure and IP Warm‑Up for an Email Service Provider
Your ESP's sending side is the MTA, the IP pools your customers share, the IPv4 addresses you lease, their reverse DNS and the warm-up that builds their reputation. This guide covers the choices we make on PowerMTA, what went wrong on the way and a warm-up schedule you can copy.
By Cem Hurturk, founder of Octeth Updated 13 min read
- PowerMTA configuration
- DKIM keys
- Reverse DNS
- Log routing
On this page
Octeth builds and queues the mail. The MTA delivers it from your IPs. Everything a mailbox provider judges about your IPs is set up in this stage: which MTA you run, which customers share which IPs, where the IPs come from, their reverse DNS and how fast you raise volume on them. The MTA runs on its own server, set up in Stage 3, server infrastructure, and talks to the Octeth install from Stage 4, installing Octeth. Sender domain records (SPF, DKIM, DMARC) are stage 6, sender domain authentication.
Choose the MTA
Octeth hands mail to the MTA over SMTP. A delivery server in Octeth is an SMTP host, port and login, so any MTA that accepts authenticated SMTP can deliver for it. Our runbooks cover PowerMTA.
| MTA | License | Price on the vendor’s site |
|---|---|---|
| PowerMTA (Bird) | Commercial. Each license is limited by server count or by volume. With a server license, Standard allows at most three outbound connections and Enterprise has no connection limit. The three-connection cap does not apply to volume licenses. Annual licenses include email support and updates. | None. The pricing page offers a custom license through sales. |
| KumoMTA | Apache 2.0. No license keys and no per-server or per-message fees. Enterprise support is an optional subscription. | None for support on the pages we checked. The software is free. |
| Postfix | IBM Public License 1.0 or Eclipse Public License 2.0, at your choice. Free. | Free. |
Sources: Bird PowerMTA pricing, Bird product terms, KumoMTA, Postfix 3.3.0 announcement.
Whichever you pick, an ESP needs these from the MTA:
- Routing by tenant. Each plan or customer sends from IPs you assign, and the sender cannot override that.
- Rules per mailbox provider. Cap connections per provider, set message rates per domain and map provider replies to actions: back off on rate limits, bounce the queue on a reputation block, bounce one recipient on an authentication failure.
- DKIM signing with a key and selector per platform on the MTA, used on every IP that platform sends from.
- An event log you can ship back to Octeth for bounces, deliveries and complaints.
PowerMTA covers all four, and Octeth reads its accounting log. If you choose KumoMTA or Postfix, plan how bounce and complaint events get back into Octeth before you send.
How to design IP pools for an ESP
Everyone on a shared IP shares its reputation. Google says it directly: the activity of any sender on a shared IP affects the reputation of all senders on it (Gmail sender guidelines). Pool design decides whose mail can hurt whose.
Split by plan first
Give each plan its own IP. Put the free plan, which is usually the signup default, on an IP it shares only with the suspended and cancelled groups, and give each paid plan a separate IP. A new free account that sends badly then hurts only other free accounts.
Customers who get dedicated IPs send from their own IPs only. Stage 1, business planning, covers which customers get them and how many IPs to plan for each.
New, trusted and quarantine
- New. From Octeth 6.0.1, new signups are created as Untrusted. Their campaigns wait for staff approval and their API sends are refused until staff mark the account Trusted. Review every new account before you mark it Trusted. Keep new accounts on the same IPs as your lowest plan, never on paid or dedicated IPs.
- Trusted. An account moves to its plan’s IPs when it upgrades, because the plan’s user group carries the delivery server. Stage 8, abuse prevention, sets the bounce and complaint limits that decide whether it may stay there, and its trust levels diagram shows which IP each level sends from.
- Quarantine. When one sender drives blocks on a shared IP, find that sender in the MTA log and move it off the shared pool. Our PowerMTA runbook does this before it throttles or pauses an IP, because on a shared IP the fix is usually one customer. Quarantine contains damage. It is not a place to keep sending the same mail, as the pitfalls below show. PowerMTA can also send a sender’s mail to a discard target, which stops delivery without changing the account in Octeth.
Route in the MTA, not by header
Fix the IP for each plan in the MTA. Give each plan its own SMTP user whose source has a fixed default IP or pool and does not accept a header that picks one. In Octeth, create one delivery server per plan with that user’s login, and assign it in the plan’s user group. A pattern for PowerMTA, with documentation IPs:
<virtual-mta growth-1>
smtp-source-host 192.0.2.11 mta11.example.com
domain-key esp1,*,/etc/pmta/dkim/esp1.pem
</virtual-mta>
<virtual-mta growth-2>
smtp-source-host 192.0.2.12 mta12.example.com
domain-key esp1,*,/etc/pmta/dkim/esp1.pem
</virtual-mta>
<virtual-mta-pool growth-pool>
virtual-mta growth-1
virtual-mta growth-2
</virtual-mta-pool>
<smtp-user plan-growth>
password <from your secrets store>
source {source-plan-growth}
</smtp-user>
<source {source-plan-growth}>
smtp-service yes
require-auth true
always-allow-relaying yes
process-x-virtual-mta no
process-x-job yes
default-virtual-mta growth-pool
</source>
PowerMTA uses the IPs in a pool in turn. With process-x-virtual-mta no, an X-Virtual-MTA header in the message is ignored. process-x-job yes makes PowerMTA record the X-Job header as the job ID, which the bounce route below depends on.
Every IP of the platform signs with the same selector, esp1 here, because each customer publishes one DKIM record for that selector in Stage 6, sender domain authentication. A selector per plan would fail DKIM for every plan but one. Signing also needs dkim-sign yes in <domain *>. Leave dkim-identity at its default, sender-or-from, so d= is the From domain.
Dedicated IP customers
Route a dedicated IP customer with the per-account delivery server override (per channel from 6.0.7) in Octeth (staff area, the customer’s settings). Do not clone its plan’s user group for it. Billing checks group membership every hour and moves an account in a group that is not linked to its plan back to the plan’s default group. Octeth decides the server in this order, each step overriding the last: the user group’s server for the channel, then a recipient MX route, then the per-account override, then segment routing for campaigns.
Separate transactional from marketing mail
Octeth assigns delivery servers per user group and per channel: marketing campaigns, transactional email and autoresponders each have their own setting. Point the transactional channel of every plan at its own delivery server, and that server at its own IP. Then a campaign that draws complaints does not slow password resets.
Tune the transactional IP for speed. Give that mail top queue priority, no message rate cap, a retry interval well below PowerMTA’s defaults (10 minutes, one hour in backoff) and a bounce window of hours instead of the default four and a half days, so a stuck one-time code fails fast instead of arriving hours late.
Set the split at the user group. Before 6.0.7 the per-account override moves all three channels together. From 6.0.7 staff can set it per channel.
Keep your own system mail apart too. Send it to the MTA through its own SMTP login, not a plan’s delivery server, and from an IP that no customer plan uses.
SMTP relay and API for your customers
Customers who send transactional mail from their own apps use Octeth’s Email Gateway: an HTTP API and an SMTP relay, with credentials per verified sender domain. The relay listens on ports 25 and 587. You can map 2525 to the 587 listener for networks that block 587. The Email Gateway API reference has the commands.
Set three things before customers connect:
- Trust gate. From Octeth 6.0.1, a setting refuses gateway sends from accounts that are not Trusted, and it is on by default. Keep it on. Mark an account Trusted only after review, as in the pool rules above.
- TLS on the relay. Give the relay a certificate. From 6.0.5 the relay requires STARTTLS on 587 and 2525 before login, and enforces it only when a certificate is configured.
- Connections per IP. Installs made on 6.0.0 or 6.0.1 have a limit of one connection per client IP, and upgrades keep it. Fresh installs from 6.0.2 ship 10. With a limit of 1, a client that opens parallel connections, or a monitor that probes two ports at once, gets refused. Check the value in the relay’s
config.json.
You can launch without the customer SMTP relay.
Source IPv4 addresses
New IPv4 space is no longer handed out. RIPE NCC made its last allocation from its free pool on 25 November 2019, and now offers members one /24 from a waiting list when addresses come back (RIPE NCC). So an ESP leases IPs from a hosting provider or a broker. A leased /24 is a monthly cost per block. Budget it in the unit economics from Stage 1, business planning.
Check the neighbourhood
Your IPs inherit the history of the range around them. One blocklist, ivmSIP/24, lists whole ranges where spam-sending patterns were found (invaluement). When it lists a /24, every IP in that block is affected. Before you send from a leased range:
- Look up each IP and its /24 on the major blocklists. Spamhaus answers only resolvers that are not public or open, and returns 127.255.255.254 otherwise, so check the return code. Invaluement, Proofpoint and Cloudmark have web lookups only.
- Confirm you can set reverse DNS for every IP, and how. Some providers make PTR changes only through their support desk.
- Ask the provider about the range’s past use and its abuse process.
- Fix the senders before you move to a clean range. For a listed /24, fix authentication and stop the worst accounts first, and move to the clean range last. Moving first would only burn the second range.
If someone else leases the IPs for you, agree who keeps them when you part. In our managed ESP, IPs we lease stay with us.
Port 25
The MTA must connect out to mailbox providers on port 25. Many clouds block it by default:
- Hetzner Cloud blocks outbound ports 25 and 465 and unblocks them on request after the account is a month old and has paid its first invoice (Hetzner).
- Amazon EC2 restricts port 25 until you submit a request with your use case (AWS).
- Vultr blocks it on new instances, as Stage 3, server infrastructure, notes.
Check the policy before you order the MTA server, then test the opposite direction too: the MTA’s port 25 must answer for its own bounce domains. The open port checker tests a port on any IP from outside. For other hosts, see VPS providers and their port 25 rules.
Set reverse DNS
Gmail requires every sending IP to have a PTR record that resolves to a hostname, and that hostname to have an A record that resolves back to the same IP (Gmail sender guidelines). Use the same name as the MTA’s HELO for that IP.
For each sending IP:
; reverse zone, set through the IP provider
11.2.0.192.in-addr.arpa. IN PTR mta11.example.com.
; forward zone
mta11.example.com. IN A 192.0.2.11
The matching PowerMTA entry uses the same name as its HELO:
<virtual-mta growth-1>
smtp-source-host 192.0.2.11 mta11.example.com
</virtual-mta>
Set the PTR before the IP sends its first message. The names can sit under the range’s own reverse DNS domain rather than under the domain that customers’ sender domains point to, as long as PTR, A record and HELO match. Look up the PTR with dig -x and the A record with dig before the first send.
Get bounces back into Octeth
Octeth suppresses an address only when the bounce reaches it. With PowerMTA, the bounces come from the accounting log:
- Each message carries a job ID with the account’s user ID in an
X-Jobheader. Octeth does not add this header for you. Add a custom header such asx-job: <prefix>-u%User:UserID%on the staff Email delivery screen, for all email types. Give each platform on the MTA its own prefix, and make sure the MTA source processes the header (process-x-job yesabove). - A log shipper, such as Vector, tails the accounting files and routes rows by that prefix.
- The shipper posts hard bounces and gateway events to the Octeth bounce webhook, signed with the webhook’s signature header.
Count on synchronous bounces, the ones the receiving server returns during delivery. With verified sender domains, the return path sits under the customer’s domain, and an MTA that accepts inbound mail only for its own relay domains refuses bounces sent back later to those return paths. Do not open the MTA to relay for every domain to catch them. That makes it an open relay. If you need delayed bounces, run a separate inbound MX for the return-path hosts.
Then prove it daily. Have the app server send one message a day to a non-existent address under a domain you control, wait for Octeth to mark it bounced, and alert when that stops. After each Octeth upgrade, check that the MTA’s posts still return 200.
How to warm up new IPs
Start at 100 emails per IP on day one and raise volume by the same factor each sending day for 30 sending days, mailing your most engaged contacts first. A new IP has no sending history. Our PowerMTA notes record Microsoft as the provider hardest on new IPs. Run the warm-up through user group sending limits in Octeth and through who you mail, and agree the plan with the customer whose mail fills the IP.
The schedule
This is the Standard profile of our IP warm-up plan generator, with weekends skipped. The example target is 80,000 a day per IP. Enter your own IPs, target and provider mix in the generator for your plan and its per-provider split.
- Days 1 to 6: 100 to 317 Opened or clicked in the last 30 days
- Days 7 to 15: 399 to 2,521 Engaged in the last 60 days
- Days 16 to 26: 3,174 to 31,817 Engaged in the last 90 days
- Days 27 to 30: 40,065 to 80,000 Full engaged list
Show the day-by-day numbers
| Sending day | Per IP | Who to mail |
|---|---|---|
| 1 | 100 | Opened or clicked in the last 30 days |
| 2 | 126 | Same |
| 3 | 159 | Same |
| 4 | 200 | Same |
| 5 | 251 | Same |
| 6 | 317 | Same |
| 7 | 399 | Engaged in the last 60 days |
| 8 | 502 | Same |
| 9 | 632 | Same |
| 10 | 796 | Same |
| 11 | 1,002 | Same |
| 12 | 1,262 | Same |
| 13 | 1,590 | Same |
| 14 | 2,002 | Same |
| 15 | 2,521 | Same |
| 16 | 3,174 | Engaged in the last 90 days |
| 17 | 3,997 | Same |
| 18 | 5,033 | Same |
| 19 | 6,337 | Same |
| 20 | 7,980 | Same |
| 21 | 10,049 | Same |
| 22 | 12,654 | Same |
| 23 | 15,935 | Same |
| 24 | 20,066 | Same |
| 25 | 25,267 | Same |
| 26 | 31,817 | Same |
| 27 | 40,065 | Full engaged list |
| 28 | 50,452 | Same |
| 29 | 63,531 | Same |
| 30 | 80,000 | Same |
Each day is about 1.26 times the day before. Hold at the target for five to seven days before you add any unengaged contacts, and add them in small separate sends.
How long does IP warm-up take?
Plan on four to six weeks for a new IP. The schedule above takes 30 sending days to reach 80,000 emails a day per IP, which is about six calendar weeks with weekends skipped.
When to stop climbing
The generator holds the ramp flat when the complaint rate passes 0.1% or the hard bounce rate passes 2%, and holds for three to five days at the last good volume when deferrals climb. Gmail asks senders to keep the spam rate in Postmaster Tools below 0.10% and never reach 0.30% (Gmail sender guidelines). Expect Microsoft to throttle a new IP first. Watch deferrals per provider in the MTA log every day of the warm-up, and decide in advance how long deferred mail may wait before it bounces.
New customers on warm IPs
Most new customers join IPs that are already warm. Their risk is a list your IPs have never mailed. Keep them on the new-account IPs, start them in a group with a low daily limit and move them to the plan’s group as their bounce and complaint rates hold. Stage 8, abuse prevention, sets those limits.
Before you move on
- You have picked an MTA and know what its license costs and limits.
- Each plan, and each dedicated IP customer, maps to a fixed IP or pool in the MTA.
- No SMTP user that Octeth or a customer can reach lets a message choose its own IP.
- Free and unvetted accounts send from IPs that no paying plan uses.
- Transactional mail has its own delivery server and IP, assigned in every plan's user group.
- You checked every leased IP and its /24 on the blocklists before the first send.
- Your MTA host allows outbound port 25, and you tested it.
- Every sending IP has a PTR record, an A record that points back to it, and the same name as its HELO.
- A test bounce reaches Octeth, and a daily canary alerts you when bounces stop.
- Customers who use the API or SMTP relay are marked Trusted only after review.
- The SMTP relay has a TLS certificate and a per-IP connection limit above 1.
- Each new IP has a warm-up plan, and you know which engaged segment feeds each phase.
What we've seen go wrong
Customers could pick any IP on the MTA
When one platform was first wired to a shared PowerMTA server, every plan used one SMTP user with header-based virtual MTA selection on. Octeth forwards headers that customers set on gateway sends to the MTA, so with that selection on, a customer could pick any virtual MTA on the box, including those of other senders, with an X-Virtual-MTA header. Leave process-x-virtual-mta off for every SMTP user Octeth uses. Give each plan its own SMTP user with a fixed IP and header selection off, then confirm that a forged header is ignored.
A blocklisting that covered the whole range
On a shared MTA whose sending IPs sat in one range, a blocklist that lists ranges listed the whole range. Mail behind that list bounced from every IP in the range, so rotating senders to other IPs in the block changed nothing. The cause was the senders: most of the top sending domains failed SPF and DKIM, and the recipients looked like a scraped or bought prospect list.
A banned sender moved to a fresh IP
Microsoft banned the IP of one account on a shared MTA. The account was moved to another IP and kept sending the same mail, and Microsoft banned the second IP too. Moving a sender contains the damage only if the sending stops or changes. Otherwise it spends one more IP.
Bounces that went to the wrong platform
When a second platform started sending through a PowerMTA server that another platform already used, the log shipper sent every row it did not recognise to the first platform. The second platform's bounces never reached it, so it would have kept mailing dead addresses. Give each platform its own job prefix in the X-Job header and its own route to Octeth's bounce webhook.
Bounce rows lost after log rotation
A daily canary bounce was never posted to Octeth. PowerMTA had deleted an old accounting file and the new file reused its inode, so the log shipper, Vector, resumed at the old file's offset and skipped the rows. We switched Vector to fingerprint the first data row of each file. Without the canary this would have gone unnoticed.
A retry storm during a Microsoft throttle
Microsoft 365 throttled one IP on a shared MTA for two days. The MTA's global settings retried deferred mail often and bounced it within hours, so it made millions of retries a day and most of the deferred mail expired as bounces before the throttle lifted. Keep global settings for bulk mail and give the transactional IP its own retry and expiry settings. Decide both per IP on purpose, and watch deferrals during every warm-up.
The MTA could not reach the app
When we set up bounce posts from the MTA to the app, the MTA's network could not reach the Cloudflare addresses the app domain resolved to. One workaround is to pin Cloudflare addresses that work in the MTA's hosts file, and remove them once the provider fixes its routing. Test the bounce webhook from the MTA itself, not from your desk.
A DKIM key that changed with the IP
On one PowerMTA server, most IPs signed with one DKIM key and a few older IP entries had no key of their own and fell back to a different global key. A sender moved between the two sets would sign with a different key without any warning. Give every IP entry its key explicitly, and give each platform that shares the MTA its own selector.
Related tools
We build ESPs
Rather have us build this?
We set up the PowerMTA server and your sender IPs, and plan the warm-up before full volume.