Do you have any sample data to confirm this? If you can't share sample data, can you let us know the average tuple size?
I suspect the writes are updates to jsonb columns. I can correlate bursts of writes of this form to the memory spikes: UPDATE suspect_table
SET jsonb_column = jsonb_column || $1,
updated_at = $2
WHERE ...
The update being added in $1 is typically a single new field. The jsonb column is flat string key/value pairs, e.g. lots of {"key": "value", ...}.
The average size of the whole tuple in the suspect table is ~800 bytes (based on 10000 random samples), of which the jsonb column is 80%.
I have been trying to break into a walsender to inspect some tuple bufs directly and compare the ChangeSize vs GetTupleBuf size as you suggest, but it's proving a little tricky - I'll let you know if I have any luck here.