My setup:

RB751G-2HnD
Draytek Vigor 120 PPPoE ADSL modem (UK model) to UK ISP running PPPoA

I purchased these about 3 years ago.

I have been upgrading ROS every few months.

For the first 2 years PPPoE achieved uptimes measured in weeks or months, up to
and including ROS 6.18

For the last year, I have noticed many PPPoE disconnections followed by immediate
reconnections, running ROS post 6.18. This is according to the logs from
my ISP to which I have access. Typical uptimes are a few minutes to several
hours. The reason given for the disconnection in the logs is usually "port error"
or "user request".

These reconnections screw up Youtube videos or BBC iPlayer downloads or
any other bulk download. This morning a download from iPlayer kept
on stopping. The ROS version was 6.32.3. I downgraded to 6.18 and
iPlayer worked again. I did not change any other configuration.

I've read the changelogs for ROS and I see nothing to indicate the code
for PPPoE has chaned, so it is a bit of a mystery why PPPoE is
so unstable post 6.18

I've kept a diary of ROS installs and ISP logs and there is definite
correlation between PPPoE instability and ROS version. I have
run 6.18 for roughly half the time in the past year.

Upgrade your router to 6.32.3. Run this command: "/system logging add topics=pppoe,debug".

After that wait until you see this problem. Now check that connection debug messages are visible in Log and generate supout file. Send this file to support@mikrotik.com. We will take a look on it.

There is a problem with my ethernet link. PPPoE is OK

More details to follow after more testing.

Here is the final report as promised.

The fault is now fixed. There is nothing wrong with the RB751.

I replaced the Vigor 120.

Its ethernet interface was faulty. There was severe packet loss on ping, probably due to many tx-collisions. The PPPoE link did not stay up for more than about 5 minutes due to this.

It seems that the fault had slowly been getting worse over the last few weeks.

Its ethernet interface was faulty. There was severe packet loss on ping, probably due to many tx-collisions.

Did you check if maybe one of the ends was configured with fixed parameters?
This is a common mistake on UTP Ethernet. There is a config option somewhere to select between AUTO and
manual setting of 10/100(/1000) and FD/HD (plus a flowcontrol option).
Some people think "it will never hurt to set this to 100 FD as that is what I want", but actually it DOES hurt to do that, when you do it only at one end and not at the other.

Why? because when an ethernet interface has to go through AUTO configuration it has to talk to the other side to negotiate the settings, and when that fails it falls back to a default config which unfortunately is Half Duplex (HD) as that was the common situation in the days this was invented.
So when you set one end to 100 FD and leave the other end at AUTO, that end will select 100 HD and indeed you will have many collisions.
(the interface will not reply to AUTO configuration requests when it is set to a fixed configuration)

So you need to have both ends at AUTO or both ends at the same fixed setting.
It is a very common pitfall and it results in many interesting observations, including what you experienced, but also poor transfer rates, often depending on the direction and the way it is tested.

reply to pe1chl

The Vigor 120 does not have any configuration for its ethernet port, at least none that I can find. I reset the Vigor to its default config, just in case, although I have never changed anything. It's supposed to work out of the box. I know about its access via Telnet and its Web interface.

I plugged the Vigor into the ethernet port on my Apple iMac, and ifconfig on the iMac reported 10M/half.

I tested the new Vigor on the iMac and it reported 100M/full.

I tried the old Vigor on the RB 751 and it was reported as advertising 10M/half. I also tried disabling auto-negotiation and running it at 10M/half. There were rx-fcs and tx-collision errors.

I then tried the new Vigor on the RB 751 and it was now advertising 10/100/full/half. The new Vigor has now been running for 21 hours without PPPoE disconnecting. There are no rx or tx errors on the interface.

I suspect the old Vigor has been slowly dying over the last year, although it seemed to work a lot better with ROS 6.18 than later versions. Maybe ROS 6.18 was more tolerant of ethernet errors. Only a couple of days ago I transferred 4GB using ROS 6.32.3. The Vigor seems to have died almost completely now in the last day or two.

Ok...
The ethernet port can be set via commandline (telnet) using the port command.
I am using a Vigor 130 but I think it is the same in the Vigor 120.
E.g.: port 1 100F

However, when something is bad rather than a parameter has been set wrongly, of course it will not fix it.
Slowly appearing problems after longer use usually indicate bad capacitors, normally in the power supply but can also be local decoupling caps in the circuit of e.g. the ethernet port.
Easy to repair when you are in electronics.

Good to hear that it is now OK!