The questions below come up constantly, and most of them have a precise answer that depends on which engine you are using.
Almost always a flavour mismatch. regex101 defaults to a specific flavour and you have to select PCRE2, Python, JavaScript, Go or Java from the sidebar. The differences that bite: JavaScript historically lacked lookbehind (it has had it since ES2018 but older environments do not), Go's RE2 has no lookaround or backreferences at all, and Java requires every backslash to be doubled in string literals so \\d in source means \d in the pattern. Set the flavour to match your runtime before debugging anything else.
PCRE, .NET, Java, Python's re and modern JavaScript all support it. .NET and JavaScript allow variable-length lookbehind; PCRE and Python require fixed length, so (?<=ab|cde) fails in Python because the alternatives differ in length. Python's third-party regex module lifts that restriction and is a drop-in replacement. Go and Rust have no lookbehind by design.
Nested quantifiers over overlapping character classes cause the engine to explore exponentially many ways to split the input. The signature is a pattern that runs instantly on short strings and hangs on slightly longer ones, always on input that fails to match. ^(a+)+$ against 30 a's followed by a b will run effectively forever.
To detect it, use the step counter in regex101's debugger — anything in the tens of thousands of steps for a short subject is a warning. To fix it, remove the nesting (^a+$), use a possessive quantifier (a++) or an atomic group ((?>a+)), or make the inner class non-overlapping.
Yes, for anything you will read again. (?P<year>\d{4}) in Python, (?<year>\d{4}) in JavaScript, .NET and PCRE. Access by name rather than index, so inserting a group later does not silently break every downstream reference. For groups you only need for alternation or grouping, use non-capturing (?:...) — it is faster and keeps the capture indices meaningful.
Verbose or extended mode. With re.X in Python or the x flag in PCRE, whitespace and # comments inside the pattern are ignored:
pattern = re.compile(r"""
^(?P<proto>https?):// # scheme
(?P<host>[^/:?\#]+) # hostname
(?::(?P<port>\d+))? # optional port
""", re.X)Remember to escape literal spaces and # in verbose mode. JavaScript has no equivalent flag; build the pattern from an array of strings and join it, or use a tagged template.
Not with classic regex — nesting is not a regular language. PCRE has recursion ((?R)) and .NET has balancing groups, both of which can handle balanced brackets, but the result is unreadable and slow. For HTML, JSON, YAML or source code, use a parser. Regex on HTML works right up until an attribute contains a >, which it eventually will.
By default \w and \d mean different things per engine. In Python 3 they are Unicode-aware by default; in JavaScript they are ASCII-only unless you use the u or v flag; in PCRE you need /u and often (*UTF). Unicode property escapes — \p{L} for any letter, \p{Script=Greek}, \p{Emoji} — are supported in Python's regex module, PCRE, .NET, Java and JavaScript with the u flag. Also note that . matches a code point, not a grapheme, so an emoji with a skin-tone modifier counts as several.
regex101 for the step debugger and the flavour switcher, which are the two features that matter most. RegExr for a cleaner learning interface. Debuggex for railroad diagrams when a pattern is too tangled to read. On the command line, rg uses Rust's linear-time engine so it cannot hang, and grep -P gives you PCRE when you need lookaround. Whatever you use to build the pattern, write unit tests for it — a regex with five test cases pinned down is maintainable and one without them is a liability.