This depends on your modulation choice and approach to carrier recovery. But consider this for a specific case of a raised cosine pulse shaped signal:
Observe (using the eye diagram would be simple) the significant symbol to symbol jitter everywhere except in close proximity to the symbol decisions — for this case I would opt to use 1 sample per symbol (the decisions) to estimate symbol to symbol phase since those samples have the least jitter (this would be the samples after all equalization, timing recovery and matched filtering is complete, just prior to decision where the complex phase can still be measured between two successive symbols).
This approach works well if you are using the Gardner TED for symbol clock recovery since it can function well with relatively large carrier offsets.
You can evaluate your specific waveform with the same considerations to determine what approach would have the minimum estimation error (and not the error on any given sample but the rms error within your carrier tracking loop bandwidth).