Comment by eugenekay
1 day ago
Throughout the late 90s, “Mail.com” provided white-label SMTP services for a lot of businesses, and was one of the early major “free email” providers. Each Free user had a storage limit of something like 10MB, which is plenty in an era before HTML email and attachments were commonplace. There were racks upon racks of SCSI disks from various vendors for the backend - but the front end was all standard Sendmail, running on Solaris servers.
Anyway, here’s the front end SMTP servers in 1999, then in-service at 25 Broadway, NYC. I am not sure exactly which model these were, but they were BIG Iron! https://kashpureff.org/album/1999/1999-08-07/M0000002.jpg
Those look like E5500 or E6500 cabinets (hard to tell from the angle).
A few months earlier there was only a single “Enterprise 6000” cabinet: https://kashpureff.org/album/1999/1999-05-30/M0000024.jpg
I worked at a competing white-label email provider in the 90s and even then it seemed obvious that running SMTP on a Sun Enterprise was a mistake. You're not gaining anything from its multiuser single-system scalability. I guess it stands as an early example of pets/cattle debate. My company was firmly on the cattle side.
I was just the Teenage intern responsible for doing the PDU Cabling every time a new rack was added, since nobody on the Network or Software Engineering teams could fit into the crawl spaces without disassembling the entire raised-floor.
I do know that scale-out and scale-up were used for different parts of the stack. The web services were all handled by standard x86 machines running Linux - and were all netbooted in some early orchestration magic, until the day the netboot server died. I think the rationale for the large Sun systems was the amount of Memory that they could hold - so the user name and spammer databases could be held in-memory on each front end, allowing for a quick ACCEPT or DENY on each incoming message - before saving it out to a mailbox via NFS.
Makes sense, there are a lot of reasons why having some "big iron" might have been practical in that era. x86 was not a full contender for many workloads until amd64, and a lot of the shared-nothing software approaches were not really there until later.