ranweiler / dent Goto Github PK

Add a flag to the exe to support emitting TSV output with full floating point precision, with or without headers By default, this should just emit TSV for each sample summary. If a flag is set to perform significance tests, then emit only those.

Show more precision by default, maybe make it configurable, and in any case, ensure column headers are aligned with data.

Right now, it is an error to invoke dent sample.txt. Instead, it should act like dent -s < sample.txt.

Right now, if any line in a file or steam of sample data cannot be parsed as a float, dent panics. We can improve on this in a few ways:

• Add a permissive line reading option, which ignores blank or otherwise malformed lines. We could make this act like a verbosity flag. So, for example, if the long flag name were --lax and the short name were -l, -l would skip blank lines, but -ll would skip any line that can't be parsed as an f64.
• If we do find an error, print useful info to stderr, such as a fragment of the bad input. For files, this can include line number.

Right now, especially when sample data is very far from normally distributed, we fail in weird ways, including some which panic (due to our liberal first-draft usage of unwrap()). Remove the panics and make plot rendering fail nicely.

The box plots visually indicate the lower and upper quartiles of each sample, but we doesn't print their values. We should also indicate the diagram scale (e.g. the add labels for the overall min and max across samples).

Make this interact nicely with plotting, too.

Right now, we require the user to enter non-default significance levels in a very rigid format, and they must choose from a small set of such values. At a minimum, parse these levels more forgivingly, and explore moving from a table of critical values to runtime computation. However flexible we are or are not, indicate the situation in the binary tool's help text.

Right now, we always perform a t-test exactly when we have two input samples. We should add flags to disable/enable this, and perhaps perform tests on all input files when there are more than two. For example, given input files a, b, and c, we could either run and display t-tests for {a, b} and {b, c} (compare the sequential pairs), or compare every two-element subset of {a, b, c}. The former is probably more useful in practice.

If we plot > 1 samples, we should plot them on the same scale by default, so they are visually comparable. We currently do this with exactly 2 samples, when performing t-tests, but in an ad hoc way. This is probably best done when improving our corner case plot handling.

Support both single data set plots when summarizing and scaled, stacked plots when comparing file data. Include a flag in the binary, and expose in the library.

Especially when it comes to TTY plotting, a few extreme outliers can instantly blow our ratio budgets. For example, if the ratio (max - min) / (q3 - q1) is too large, we won't have the resolution to display a boxplot, because we can't make the box small enough relative to the overall size of the figure.

We can address this by providing options to do any of the following:

• Filter outliers from the input data before any processing
• Filter outliers from the input data before plotting only
• Plot outliers, but determine the outer fences by a formula like Q ± 1.5 IQR, where Q is the 1st or 3rd quartile.
• Allow different notions of "outlier"

If two samples A and B have significantly different means, compute and display the difference in those means, along with a confidence interval.

Right now, the Summary struct has a data field, and provides methods to recompute statistics on that sample data, on demand. Instead, make Summary just a read-only struct of the summary statistics without the data (which isn't editable or even readable right now), computed on construction. For now, continue to make a temporary sorted copy of the entire data set, but drop it when we return from the ctor.