feat: add FileEdit and VersionEdit

src/version.zig

@@ -0,0 +1,156 @@
+const std = @import("std");
+const encode = @import("encode.zig");
+
+const Allocator = std.mem.Allocator;
+
+// VersionSet/VersionEdit: append-only log of incremental changes to the
+// live sstable set (fileId -> level), replacing a full-state rewrite on
+// every flush/compact. Modeled on LevelDB's VersionSet: a CURRENT file
+// names the active manifest log, the log holds a sequence of edit
+// records, and recovery replays them into a map. The log is periodically
+// rotated (a single snapshot record replaces the whole history) so it
+// never grows without bound.
+
+pub const VersionError = error{
+    Corrupt,
+    Truncated,
+};
+
+pub const FileEdit = struct {
+    fileId: u64,
+    level: u32,
+};
+
+pub const VersionEdit = struct {
+    allocator: Allocator,
+    added: std.ArrayList(FileEdit),
+    deleted: std.ArrayList(u64),
+    /// Optional counter watermarks, stamped by the caller alongside a file
+    /// add/delete: the exact `Db.nextFileId`/`Db.nextSeq` value in effect
+    /// right after this edit was built. Recorded so recovery can recover
+    /// these counters from the manifest instead of relying solely on
+    /// scanning which sstable files currently happen to be on disk - a scan
+    /// alone regresses whenever a compaction fully garbage-collects a run of
+    /// tombstones into zero output files (see runMerge: no output file means
+    /// no fresh fileId is minted, and if the GC'd inputs held the highest
+    /// seq/fileId ever used, a post-restart scan would compute a lower
+    /// floor than what was really already consumed, risking seq reuse
+    /// across a restart).
+    nextFileId: ?u64 = null,
+    nextSeq: ?u64 = null,
+
+    pub fn init(allocator: Allocator) VersionEdit {
+        return .{
+            .allocator = allocator,
+            .added = std.ArrayList(FileEdit).init(allocator),
+            .deleted = std.ArrayList(u64).init(allocator),
+        };
+    }
+
+    pub fn deinit(self: *VersionEdit) void {
+        self.added.deinit();
+        self.deleted.deinit();
+    }
+
+    pub fn addFile(self: *VersionEdit, fileId: u64, level: usize) !void {
+        try self.added.append(.{ .fileId = fileId, .level = @intCast(level) });
+    }
+
+    pub fn deleteFile(self: *VersionEdit, fileId: u64) !void {
+        try self.deleted.append(fileId);
+    }
+
+    pub fn setNextFileId(self: *VersionEdit, nextFileId: u64) void {
+        self.nextFileId = nextFileId;
+    }
+
+    pub fn setNextSeq(self: *VersionEdit, nextSeq: u64) void {
+        self.nextSeq = nextSeq;
+    }
+
+    pub fn isEmpty(self: *const VersionEdit) bool {
+        return self.added.items.len == 0 and self.deleted.items.len == 0;
+    }
+
+    fn encode_(self: *const VersionEdit, out: *std.ArrayList(u8)) !void {
+        const writer = encode.ByteWriter.init(out);
+        try writer.putU32(@intCast(self.added.items.len));
+        for (self.added.items) |f| {
+            try writer.putU64(f.fileId);
+            try writer.putU32(f.level);
+        }
+        try writer.putU32(@intCast(self.deleted.items.len));
+        for (self.deleted.items) |id| try writer.putU64(id);
+        try writer.byte(if (self.nextFileId != null) 1 else 0);
+        if (self.nextFileId) |v| try writer.putU64(v);
+        try writer.byte(if (self.nextSeq != null) 1 else 0);
+        if (self.nextSeq) |v| try writer.putU64(v);
+    }
+};
+
+/// Applies an edit's added/deleted entries onto a live fileId->level map.
+pub fn applyEdit(state: *std.AutoHashMap(u64, usize), added: []const FileEdit, deleted: []const u64) !void {
+    for (deleted) |id| _ = state.remove(id);
+    for (added) |f| try state.put(f.fileId, f.level);
+}
+
+/// Recovered counter watermarks, threaded out of `VersionSet.open`'s replay
+/// so `Db.open` can fold them into its own directory-scan-based recovery
+/// (taking the max of both sources - see `VersionEdit.nextFileId`/`nextSeq`
+/// doc comment for why the scan alone is not always sufficient).
+pub const Counters = struct {
+    nextFileId: ?u64 = null,
+    nextSeq: ?u64 = null,
+};
+
+const currentName = "CURRENT";
+const manifestPrefix = "MANIFEST-";
+
+fn manifestFileName(buf: []u8, id: u64) ![]const u8 {
+    return std.fmt.bufPrint(buf, "{s}{d:0>10}", .{ manifestPrefix, id });
+}
+
+pub const VersionSet = struct {
+    allocator: Allocator,
+    dir: std.fs.Dir,
+    file: std.fs.File,
+    manifestId: u64,
+    /// Number of edit records appended to the current log since the last
+    /// rotation (used as the rotation heuristic).
+    pendingRecords: usize,
+    /// Latest counter watermarks seen across every replayed (or since
+    /// applied) edit - see `VersionEdit.nextFileId`/`nextSeq`. Carried
+    /// forward into `rotate`'s snapshot record so they survive log rotation
+    /// even though the individual edits that set them get discarded.
+    counters: Counters,
+
+    /// Opens (or bootstraps) the version set, replaying the active
+    /// manifest log into a fresh fileId->level map that the caller owns.
+    pub fn open(allocator: Allocator, dir: std.fs.Dir) !struct { set: VersionSet, state: std.AutoHashMap(u64, usize) } {
+        var state = std.AutoHashMap(u64, usize).init(allocator);
+        errdefer state.deinit();
+
+        const currentContents = dir.readFileAlloc(allocator, currentName, 256) catch |err| switch (err) {
+            error.FileNotFound => {
+                const set = try bootstrap(allocator, dir);
+                return .{ .set = set, .state = state };
+            },
+            else => return err,
+        };
+        defer allocator.free(currentContents);
+
+        const trimmed = std.mem.trim(u8, currentContents, " \t\r\n");
+        const id = parseManifestId(trimmed) orelse return VersionError.Corrupt;
+
+        const file = dir.openFile(trimmed, .{ .mode = .read_write }) catch |err| switch (err) {
+            error.FileNotFound => {
+                const set = try bootstrap(allocator, dir);
+                return .{ .set = set, .state = state };
+            },
+            else => return err,
+        };
+        errdefer file.close();
+
+        const data = try file.readToEndAlloc(allocator, std.math.maxInt(usize));
+        defer allocator.free(data);
+