annotate rust/hg-core/src/revlog/changelog.rs @ 51397:b01e7d97e167

revlog: add a Rust implementation of `headrevsdiff`

Python implementation of `headrevsdiff` can be very slow in the worst
case compared with the `heads` computation it replaces, since the
latter is done in Rust.

Even the average case of this Python implementation is still
noticeable in the profiles.

This patch makes the computation much much faster by doing it in Rust.

use std::ascii::escape_default;
use std::borrow::Cow;
use std::collections::BTreeMap;
use std::fmt::{Debug, Formatter};
use std::{iter, str};

use chrono::{DateTime, FixedOffset, NaiveDateTime};
use itertools::{Either, Itertools};

use crate::errors::HgError;
use crate::revlog::Revision;
use crate::revlog::{Node, NodePrefix};
use crate::revlog::{Revlog, RevlogEntry, RevlogError};
use crate::utils::hg_path::HgPath;
use crate::vfs::Vfs;
use crate::{Graph, GraphError, RevlogOpenOptions, UncheckedRevision};

/// A specialized `Revlog` to work with changelog data format.
pub struct Changelog {
    /// The generic `revlog` format.
    pub(crate) revlog: Revlog,
}

impl Changelog {
    /// Open the `changelog` of a repository given by its root.
    pub fn open(
        store_vfs: &Vfs,
        options: RevlogOpenOptions,
    ) -> Result<Self, HgError> {
        let revlog = Revlog::open(store_vfs, "00changelog.i", None, options)?;
        Ok(Self { revlog })
    }

    /// Return the `ChangelogRevisionData` for the given node ID.
    pub fn data_for_node(
        &self,
        node: NodePrefix,
    ) -> Result<ChangelogRevisionData, RevlogError> {
        let rev = self.revlog.rev_from_node(node)?;
        self.entry_for_checked_rev(rev)?.data()
    }

    /// Return the [`ChangelogEntry`] for the given revision number.
    pub fn entry_for_rev(
        &self,
        rev: UncheckedRevision,
    ) -> Result<ChangelogEntry, RevlogError> {
        let revlog_entry = self.revlog.get_entry(rev)?;
        Ok(ChangelogEntry { revlog_entry })
    }

    /// Same as [`Self::entry_for_rev`] for checked revisions.
    fn entry_for_checked_rev(
        &self,
        rev: Revision,
    ) -> Result<ChangelogEntry, RevlogError> {
        let revlog_entry = self.revlog.get_entry_for_checked_rev(rev)?;
        Ok(ChangelogEntry { revlog_entry })
    }

    /// Return the [`ChangelogRevisionData`] for the given revision number.
    ///
    /// This is a useful shortcut in case the caller does not need the
    /// generic revlog information (parents, hashes etc). Otherwise
    /// consider taking a [`ChangelogEntry`] with
    /// [entry_for_rev](`Self::entry_for_rev`) and doing everything from there.
    pub fn data_for_rev(
        &self,
        rev: UncheckedRevision,
    ) -> Result<ChangelogRevisionData, RevlogError> {
        self.entry_for_rev(rev)?.data()
    }

    pub fn node_from_rev(&self, rev: UncheckedRevision) -> Option<&Node> {
        self.revlog.node_from_rev(rev)
    }

    pub fn rev_from_node(
        &self,
        node: NodePrefix,
    ) -> Result<Revision, RevlogError> {
        self.revlog.rev_from_node(node)
    }
}

impl Graph for Changelog {
    fn parents(&self, rev: Revision) -> Result<[Revision; 2], GraphError> {
        self.revlog.parents(rev)
    }
}

/// A specialized `RevlogEntry` for `changelog` data format
///
/// This is a `RevlogEntry` with the added semantics that the associated
/// data should meet the requirements for `changelog`, materialized by
/// the fact that `data()` constructs a `ChangelogRevisionData`.
/// In case that promise would be broken, the `data` method returns an error.
#[derive(Clone)]
pub struct ChangelogEntry<'changelog> {
    /// Same data, as a generic `RevlogEntry`.
    pub(crate) revlog_entry: RevlogEntry<'changelog>,
}

impl<'changelog> ChangelogEntry<'changelog> {
    pub fn data<'a>(
        &'a self,
    ) -> Result<ChangelogRevisionData<'changelog>, RevlogError> {
        let bytes = self.revlog_entry.data()?;
        if bytes.is_empty() {
            Ok(ChangelogRevisionData::null())
        } else {
            Ok(ChangelogRevisionData::new(bytes).map_err(|err| {
                RevlogError::Other(HgError::CorruptedRepository(format!(
                    "Invalid changelog data for revision {}: {:?}",
                    self.revlog_entry.revision(),
                    err
                )))
            })?)
        }
    }

    /// Obtain a reference to the underlying `RevlogEntry`.
    ///
    /// This allows the caller to access the information that is common
    /// to all revlog entries: revision number, node id, parent revisions etc.
    pub fn as_revlog_entry(&self) -> &RevlogEntry {
        &self.revlog_entry
    }

    pub fn p1_entry(&self) -> Result<Option<ChangelogEntry>, RevlogError> {
        Ok(self
            .revlog_entry
            .p1_entry()?
            .map(|revlog_entry| Self { revlog_entry }))
    }

    pub fn p2_entry(&self) -> Result<Option<ChangelogEntry>, RevlogError> {
        Ok(self
            .revlog_entry
            .p2_entry()?
            .map(|revlog_entry| Self { revlog_entry }))
    }
}

/// `Changelog` entry which knows how to interpret the `changelog` data bytes.
#[derive(PartialEq)]
pub struct ChangelogRevisionData<'changelog> {
    /// The data bytes of the `changelog` entry.
    bytes: Cow<'changelog, [u8]>,
    /// The end offset for the hex manifest (not including the newline)
    manifest_end: usize,
    /// The end offset for the user+email (not including the newline)
    user_end: usize,
    /// The end offset for the timestamp+timezone+extras (not including the
    /// newline)
    timestamp_end: usize,
    /// The end offset for the file list (not including the newline)
    files_end: usize,
}

impl<'changelog> ChangelogRevisionData<'changelog> {
    fn new(bytes: Cow<'changelog, [u8]>) -> Result<Self, HgError> {
        let mut line_iter = bytes.split(|b| b == &b'\n');
        let manifest_end = line_iter
            .next()
            .expect("Empty iterator from split()?")
            .len();
        let user_slice = line_iter.next().ok_or_else(|| {
            HgError::corrupted("Changeset data truncated after manifest line")
        })?;
        let user_end = manifest_end + 1 + user_slice.len();
        let timestamp_slice = line_iter.next().ok_or_else(|| {
            HgError::corrupted("Changeset data truncated after user line")
        })?;
        let timestamp_end = user_end + 1 + timestamp_slice.len();
        let mut files_end = timestamp_end + 1;
        loop {
            let line = line_iter.next().ok_or_else(|| {
                HgError::corrupted("Changeset data truncated in files list")
            })?;
            if line.is_empty() {
                if files_end == bytes.len() {
                    // The list of files ended with a single newline (there
                    // should be two)
                    return Err(HgError::corrupted(
                        "Changeset data truncated after files list",
                    ));
                }
                files_end -= 1;
                break;
            }
            files_end += line.len() + 1;
        }

        Ok(Self {
            bytes,
            manifest_end,
            user_end,
            timestamp_end,
            files_end,
        })
    }

    fn null() -> Self {
        Self::new(Cow::Borrowed(
            b"0000000000000000000000000000000000000000\n\n0 0\n\n",
        ))
        .unwrap()
    }

    /// Return an iterator over the lines of the entry.
    pub fn lines(&self) -> impl Iterator<Item = &[u8]> {
        self.bytes.split(|b| b == &b'\n')
    }

    /// Return the node id of the `manifest` referenced by this `changelog`
    /// entry.
    pub fn manifest_node(&self) -> Result<Node, HgError> {
        let manifest_node_hex = &self.bytes[..self.manifest_end];
        Node::from_hex_for_repo(manifest_node_hex)
    }

    /// The full user string (usually a name followed by an email enclosed in
    /// angle brackets)
    pub fn user(&self) -> &[u8] {
        &self.bytes[self.manifest_end + 1..self.user_end]
    }

    /// The full timestamp line (timestamp in seconds, offset in seconds, and
    /// possibly extras)
    // TODO: We should expose this in a more useful way
    pub fn timestamp_line(&self) -> &[u8] {
        &self.bytes[self.user_end + 1..self.timestamp_end]
    }

    /// Parsed timestamp.
    pub fn timestamp(&self) -> Result<DateTime<FixedOffset>, HgError> {
        parse_timestamp(self.timestamp_line())
    }

    /// Optional commit extras.
    pub fn extra(&self) -> Result<BTreeMap<String, Vec<u8>>, HgError> {
        parse_timestamp_line_extra(self.timestamp_line())
    }

    /// The files changed in this revision.
    pub fn files(&self) -> impl Iterator<Item = &HgPath> {
        if self.timestamp_end == self.files_end {
            Either::Left(iter::empty())
        } else {
            Either::Right(
                self.bytes[self.timestamp_end + 1..self.files_end]
                    .split(|b| b == &b'\n')
                    .map(HgPath::new),
            )
        }
    }

    /// The change description.
    pub fn description(&self) -> &[u8] {
        &self.bytes[self.files_end + 2..]
    }
}

impl Debug for ChangelogRevisionData<'_> {
    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("ChangelogRevisionData")
            .field("bytes", &debug_bytes(&self.bytes))
            .field("manifest", &debug_bytes(&self.bytes[..self.manifest_end]))
            .field(
                "user",
                &debug_bytes(
                    &self.bytes[self.manifest_end + 1..self.user_end],
                ),
            )
            .field(
                "timestamp",
                &debug_bytes(
                    &self.bytes[self.user_end + 1..self.timestamp_end],
                ),
            )
            .field(
                "files",
                &debug_bytes(
                    &self.bytes[self.timestamp_end + 1..self.files_end],
                ),
            )
            .field(
                "description",
                &debug_bytes(&self.bytes[self.files_end + 2..]),
            )
            .finish()
    }
}

fn debug_bytes(bytes: &[u8]) -> String {
    String::from_utf8_lossy(
        &bytes.iter().flat_map(|b| escape_default(*b)).collect_vec(),
    )
    .to_string()
}

/// Parse the raw bytes of the timestamp line from a changelog entry.
///
/// According to the documentation in `hg help dates` and the
/// implementation in `changelog.py`, the format of the timestamp line
/// is `time tz extra\n` where:
///
/// - `time` is an ASCII-encoded signed int or float denoting a UTC timestamp
///   as seconds since the UNIX epoch.
///
/// - `tz` is the timezone offset as an ASCII-encoded signed integer denoting
///   seconds WEST of UTC (so negative for timezones east of UTC, which is the
///   opposite of the sign in ISO 8601 timestamps).
///
/// - `extra` is an optional set of NUL-delimited key-value pairs, with the key
///   and value in each pair separated by an ASCII colon. Keys are limited to
///   ASCII letters, digits, hyphens, and underscores, whereas values can be
///   arbitrary bytes.
fn parse_timestamp(
    timestamp_line: &[u8],
) -> Result<DateTime<FixedOffset>, HgError> {
    let mut parts = timestamp_line.splitn(3, |c| *c == b' ');

    let timestamp_bytes = parts
        .next()
        .ok_or_else(|| HgError::corrupted("missing timestamp"))?;
    let timestamp_str = str::from_utf8(timestamp_bytes).map_err(|e| {
        HgError::corrupted(format!("timestamp is not valid UTF-8: {e}"))
    })?;
    let timestamp_utc = timestamp_str
        .parse()
        .map_err(|e| {
            HgError::corrupted(format!("failed to parse timestamp: {e}"))
        })
        .and_then(|secs| {
            NaiveDateTime::from_timestamp_opt(secs, 0).ok_or_else(|| {
                HgError::corrupted(format!(
                    "integer timestamp out of valid range: {secs}"
                ))
            })
        })
        // Attempt to parse the timestamp as a float if we can't parse
        // it as an int. It doesn't seem like float timestamps are actually
        // used in practice, but the Python code supports them.
        .or_else(|_| parse_float_timestamp(timestamp_str))?;

    let timezone_bytes = parts
        .next()
        .ok_or_else(|| HgError::corrupted("missing timezone"))?;
    let timezone_secs: i32 = str::from_utf8(timezone_bytes)
        .map_err(|e| {
            HgError::corrupted(format!("timezone is not valid UTF-8: {e}"))
        })?
        .parse()
        .map_err(|e| {
            HgError::corrupted(format!("timezone is not an integer: {e}"))
        })?;
    let timezone = FixedOffset::west_opt(timezone_secs)
        .ok_or_else(|| HgError::corrupted("timezone offset out of bounds"))?;

    Ok(DateTime::from_naive_utc_and_offset(timestamp_utc, timezone))
}

/// Attempt to parse the given string as floating-point timestamp, and
/// convert the result into a `chrono::NaiveDateTime`.
fn parse_float_timestamp(
    timestamp_str: &str,
) -> Result<NaiveDateTime, HgError> {
    let timestamp = timestamp_str.parse::<f64>().map_err(|e| {
        HgError::corrupted(format!("failed to parse timestamp: {e}"))
    })?;

    // To construct a `NaiveDateTime` we'll need to convert the float
    // into signed integer seconds and unsigned integer nanoseconds.
    let mut secs = timestamp.trunc() as i64;
    let mut subsecs = timestamp.fract();

    // If the timestamp is negative, we need to express the fractional
    // component as positive nanoseconds since the previous second.
    if timestamp < 0.0 {
        secs -= 1;
        subsecs += 1.0;
    }

    // This cast should be safe because the fractional component is
    // by definition less than 1.0, so this value should not exceed
    // 1 billion, which is representable as an f64 without loss of
    // precision and should fit into a u32 without overflowing.
    //
    // (Any loss of precision in the fractional component will have
    // already happened at the time of initial parsing; in general,
    // f64s are insufficiently precise to provide nanosecond-level
    // precision with present-day timestamps.)
    let nsecs = (subsecs * 1_000_000_000.0) as u32;

    NaiveDateTime::from_timestamp_opt(secs, nsecs).ok_or_else(|| {
        HgError::corrupted(format!(
            "float timestamp out of valid range: {timestamp}"
        ))
    })
}

/// Decode changeset extra fields.
///
/// Extras are null-delimited key-value pairs where the key consists of ASCII
/// alphanumeric characters plus hyphens and underscores, and the value can
/// contain arbitrary bytes.
fn decode_extra(extra: &[u8]) -> Result<BTreeMap<String, Vec<u8>>, HgError> {
    extra
        .split(|c| *c == b'\0')
        .map(|pair| {
            let pair = unescape_extra(pair);
            let mut iter = pair.splitn(2, |c| *c == b':');

            let key_bytes =
                iter.next().filter(|k| !k.is_empty()).ok_or_else(|| {
                    HgError::corrupted("empty key in changeset extras")
                })?;

            let key = str::from_utf8(key_bytes)
                .ok()
                .filter(|k| {
                    k.chars().all(|c| {
                        c.is_ascii_alphanumeric() || c == '_' || c == '-'
                    })
                })
                .ok_or_else(|| {
                    let key = String::from_utf8_lossy(key_bytes);
                    HgError::corrupted(format!(
                        "invalid key in changeset extras: {key}",
                    ))
                })?
                .to_string();

            let value = iter.next().map(Into::into).ok_or_else(|| {
                HgError::corrupted(format!(
                    "missing value for changeset extra: {key}"
                ))
            })?;

            Ok((key, value))
        })
        .collect()
}

/// Parse the extra fields from a changeset's timestamp line.
fn parse_timestamp_line_extra(
    timestamp_line: &[u8],
) -> Result<BTreeMap<String, Vec<u8>>, HgError> {
    Ok(timestamp_line
        .splitn(3, |c| *c == b' ')
        .nth(2)
        .map(decode_extra)
        .transpose()?
        .unwrap_or_default())
}

/// Decode Mercurial's escaping for changelog extras.
///
/// The `_string_escape` function in `changelog.py` only escapes 4 characters
/// (null, backslash, newline, and carriage return) so we only decode those.
///
/// The Python code also includes a workaround for decoding escaped nuls
/// that are followed by an ASCII octal digit, since Python's built-in
/// `string_escape` codec will interpret that as an escaped octal byte value.
/// That workaround is omitted here since we don't support decoding octal.
fn unescape_extra(bytes: &[u8]) -> Vec<u8> {
    let mut output = Vec::with_capacity(bytes.len());
    let mut input = bytes.iter().copied();

    while let Some(c) = input.next() {
        if c != b'\\' {
            output.push(c);
            continue;
        }

        match input.next() {
            Some(b'0') => output.push(b'\0'),
            Some(b'\\') => output.push(b'\\'),
            Some(b'n') => output.push(b'\n'),
            Some(b'r') => output.push(b'\r'),
            // The following cases should never occur in theory because any
            // backslashes in the original input should have been escaped
            // with another backslash, so it should not be possible to
            // observe an escape sequence other than the 4 above.
            Some(c) => output.extend_from_slice(&[b'\\', c]),
            None => output.push(b'\\'),
        }
    }

    output
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::vfs::Vfs;
    use crate::NULL_REVISION;
    use pretty_assertions::assert_eq;

    #[test]
    fn test_create_changelogrevisiondata_invalid() {
        // Completely empty
        assert!(ChangelogRevisionData::new(Cow::Borrowed(b"abcd")).is_err());
        // No newline after manifest
        assert!(ChangelogRevisionData::new(Cow::Borrowed(b"abcd")).is_err());
        // No newline after user
        assert!(ChangelogRevisionData::new(Cow::Borrowed(b"abcd\n")).is_err());
        // No newline after timestamp
        assert!(
            ChangelogRevisionData::new(Cow::Borrowed(b"abcd\n\n0 0")).is_err()
        );
        // Missing newline after files
        assert!(ChangelogRevisionData::new(Cow::Borrowed(
            b"abcd\n\n0 0\nfile1\nfile2"
        ))
        .is_err(),);
        // Only one newline after files
        assert!(ChangelogRevisionData::new(Cow::Borrowed(
            b"abcd\n\n0 0\nfile1\nfile2\n"
        ))
        .is_err(),);
    }

    #[test]
    fn test_create_changelogrevisiondata() {
        let data = ChangelogRevisionData::new(Cow::Borrowed(
            b"0123456789abcdef0123456789abcdef01234567
Some One <[e-mail]>
0 0
file1
file2

some
commit
message",
        ))
        .unwrap();
        assert_eq!(
            data.manifest_node().unwrap(),
            Node::from_hex("0123456789abcdef0123456789abcdef01234567")
                .unwrap()
        );
        assert_eq!(data.user(), b"Some One <[e-mail]>");
        assert_eq!(data.timestamp_line(), b"0 0");
        assert_eq!(
            data.files().collect_vec(),
            vec![HgPath::new("file1"), HgPath::new("file2")]
        );
        assert_eq!(data.description(), b"some\ncommit\nmessage");
    }

    #[test]
    fn test_data_from_rev_null() -> Result<(), RevlogError> {
        // an empty revlog will be enough for this case
        let temp = tempfile::tempdir().unwrap();
        let vfs = Vfs { base: temp.path() };
        std::fs::write(temp.path().join("foo.i"), b"").unwrap();
        let revlog =
            Revlog::open(&vfs, "foo.i", None, RevlogOpenOptions::new())
                .unwrap();

        let changelog = Changelog { revlog };
        assert_eq!(
            changelog.data_for_rev(NULL_REVISION.into())?,
            ChangelogRevisionData::null()
        );
        // same with the intermediate entry object
        assert_eq!(
            changelog.entry_for_rev(NULL_REVISION.into())?.data()?,
            ChangelogRevisionData::null()
        );
        Ok(())
    }

    #[test]
    fn test_empty_files_list() {
        assert!(ChangelogRevisionData::null()
            .files()
            .collect_vec()
            .is_empty());
    }

    #[test]
    fn test_unescape_basic() {
        // '\0', '\\', '\n', and '\r' are correctly unescaped.
        let expected = b"AAA\0BBB\\CCC\nDDD\rEEE";
        let escaped = br"AAA\0BBB\\CCC\nDDD\rEEE";
        let unescaped = unescape_extra(escaped);
        assert_eq!(&expected[..], &unescaped[..]);
    }

    #[test]
    fn test_unescape_unsupported_sequence() {
        // Other escape sequences are left unaltered.
        for c in 0u8..255 {
            match c {
                b'0' | b'\\' | b'n' | b'r' => continue,
                c => {
                    let expected = &[b'\\', c][..];
                    let unescaped = unescape_extra(expected);
                    assert_eq!(expected, &unescaped[..]);
                }
            }
        }
    }

    #[test]
    fn test_unescape_trailing_backslash() {
        // Trailing backslashes are OK.
        let expected = br"hi\";
        let unescaped = unescape_extra(expected);
        assert_eq!(&expected[..], &unescaped[..]);
    }

    #[test]
    fn test_unescape_nul_followed_by_octal() {
        // Escaped NUL chars followed by octal digits are decoded correctly.
        let expected = b"\012";
        let escaped = br"\012";
        let unescaped = unescape_extra(escaped);
        assert_eq!(&expected[..], &unescaped[..]);
    }

    #[test]
    fn test_parse_float_timestamp() {
        let test_cases = [
            // Zero should map to the UNIX epoch.
            ("0.0", "1970-01-01 00:00:00"),
            // Negative zero should be the same as positive zero.
            ("-0.0", "1970-01-01 00:00:00"),
            // Values without fractional components should work like integers.
            // (Assuming the timestamp is within the limits of f64 precision.)
            ("1115154970.0", "2005-05-03 21:16:10"),
            // We expect some loss of precision in the fractional component
            // when parsing arbitrary floating-point values.
            ("1115154970.123456789", "2005-05-03 21:16:10.123456716"),
            // But representable f64 values should parse losslessly.
            ("1115154970.123456716", "2005-05-03 21:16:10.123456716"),
            // Negative fractional components are subtracted from the epoch.
            ("-1.333", "1969-12-31 23:59:58.667"),
        ];

        for (input, expected) in test_cases {
            let res = parse_float_timestamp(input).unwrap().to_string();
            assert_eq!(res, expected);
        }
    }

    fn escape_extra(bytes: &[u8]) -> Vec<u8> {
        let mut output = Vec::with_capacity(bytes.len());

        for c in bytes.iter().copied() {
            output.extend_from_slice(match c {
                b'\0' => &b"\\0"[..],
                b'\\' => &b"\\\\"[..],
                b'\n' => &b"\\n"[..],
                b'\r' => &b"\\r"[..],
                _ => {
                    output.push(c);
                    continue;
                }
            });
        }

        output
    }

    fn encode_extra<K, V>(pairs: impl IntoIterator<Item = (K, V)>) -> Vec<u8>
    where
        K: AsRef<[u8]>,
        V: AsRef<[u8]>,
    {
        let extras = pairs.into_iter().map(|(k, v)| {
            escape_extra(&[k.as_ref(), b":", v.as_ref()].concat())
        });
        // Use fully-qualified syntax to avoid a future naming conflict with
        // the standard library: https://github.com/rust-lang/rust/issues/79524
        Itertools::intersperse(extras, b"\0".to_vec()).concat()
    }

    #[test]
    fn test_decode_extra() {
        let extra = [
            ("branch".into(), b"default".to_vec()),
            ("key-with-hyphens".into(), b"value1".to_vec()),
            ("key_with_underscores".into(), b"value2".to_vec()),
            ("empty-value".into(), b"".to_vec()),
            ("binary-value".into(), (0u8..=255).collect::<Vec<_>>()),
        ]
        .into_iter()
        .collect::<BTreeMap<String, Vec<u8>>>();

        let encoded = encode_extra(&extra);
        let decoded = decode_extra(&encoded).unwrap();

        assert_eq!(extra, decoded);
    }

    #[test]
    fn test_corrupt_extra() {
        let test_cases = [
            (&b""[..], "empty input"),
            (&b"\0"[..], "unexpected null byte"),
            (&b":empty-key"[..], "empty key"),
            (&b"\0leading-null:"[..], "leading null"),
            (&b"trailing-null:\0"[..], "trailing null"),
            (&b"missing-value"[..], "missing value"),
            (&b"$!@# non-alphanum-key:"[..], "non-alphanumeric key"),
            (&b"\xF0\x9F\xA6\x80 non-ascii-key:"[..], "non-ASCII key"),
        ];

        for (extra, msg) in test_cases {
            assert!(
                decode_extra(&extra).is_err(),
                "corrupt extra should have failed to parse: {}",
                msg
            );
        }
    }

    #[test]
    fn test_parse_timestamp_line() {
        let extra = [
            ("branch".into(), b"default".to_vec()),
            ("key-with-hyphens".into(), b"value1".to_vec()),
            ("key_with_underscores".into(), b"value2".to_vec()),
            ("empty-value".into(), b"".to_vec()),
            ("binary-value".into(), (0u8..=255).collect::<Vec<_>>()),
        ]
        .into_iter()
        .collect::<BTreeMap<String, Vec<u8>>>();

        let mut line: Vec<u8> = b"1115154970 28800 ".to_vec();
        line.extend_from_slice(&encode_extra(&extra));

        let timestamp = parse_timestamp(&line).unwrap();
        assert_eq!(&timestamp.to_rfc3339(), "2005-05-03T13:16:10-08:00");

        let parsed_extra = parse_timestamp_line_extra(&line).unwrap();
        assert_eq!(extra, parsed_extra);
    }
}