<div dir="ltr">On Mon, Jul 6, 2015 at 6:28 PM Gael Varoquaux <<a href="mailto:gael.varoquaux@normalesup.org">gael.varoquaux@normalesup.org</a>> wrote:<div><br><div class="gmail_quote"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">You cannot do numerics without being aware of dtype and overflow.<br>
Seriously.<br></blockquote><div><br></div><div>When 75% of your files are natively stored as floats, but for some reason the 25% of them are shorts, it's remarkably easy to not notice that Something Bad has quietly happened, until you go and look at all the output files and say "wait, what?"</div><div><br></div><div>Yes, I know that situation is not supposed to happen (and I could have avoided it by running a data type report before doing computation), but there are delightful surprises in datasets. Even when you're aware of these things, they're still lurking with their sharp pointy teeth.</div><div><br></div><div>The ultimate answer to "what is the least surprising behavior?" is complicated and probably depends on your normal use cases. For me, it's default conversion to float.</div><div><br></div><div>-n<br></div></div></div></div>