kv-index 1.0.0

use std::{
    cmp::Reverse,
    collections::{BinaryHeap, HashMap},
    hash::{BuildHasherDefault, Hasher},
    sync::{
        atomic::{AtomicU64, Ordering},
        Arc, Weak,
    },
};

use dashmap::{mapref::entry::Entry, DashMap};
use once_cell::sync::Lazy;
use parking_lot::RwLock;
use tracing::debug;

use super::{
    common::{MatchResult, TenantId},
    RadixTree,
};

type NodeRef = Arc<Node>;

/// Shard counts for DashMaps to balance concurrency vs allocation overhead.
/// Default DashMap uses num_cpus * 4 shards (e.g., 256 on 64-core machines).
///
/// Root node uses higher shard count since ALL requests pass through it.
/// Other nodes use lower count as traffic diverges through the tree.
///
/// This reduces memory by ~90% vs default while maintaining good concurrency.
const ROOT_SHARD_COUNT: usize = 32;
const NODE_SHARD_COUNT: usize = 8;

/// Create a children DashMap for non-root nodes
#[inline]
fn new_children_map() -> DashMap<char, NodeRef, CharHasherBuilder> {
    DashMap::with_hasher_and_shard_amount(CharHasherBuilder::default(), NODE_SHARD_COUNT)
}

/// Create a tenant access time DashMap for non-root nodes
#[inline]
fn new_tenant_map() -> DashMap<TenantId, u64> {
    DashMap::with_shard_amount(NODE_SHARD_COUNT)
}

/// Result of a prefix match operation, including char counts to avoid recomputation.
#[derive(Debug, Clone)]
pub struct PrefixMatchResult {
    /// The tenant that owns the matched prefix (zero-copy)
    pub tenant: TenantId,
    /// Number of characters matched
    pub matched_char_count: usize,
    /// Total number of characters in the input text
    pub input_char_count: usize,
}

impl MatchResult for PrefixMatchResult {
    fn tenant(&self) -> &TenantId {
        &self.tenant
    }

    fn matched_count(&self) -> usize {
        self.matched_char_count
    }

    fn input_count(&self) -> usize {
        self.input_char_count
    }
}

/// A fast identity hasher for single-character keys (used in children DashMap).
/// Since chars have good distribution already, we use identity hashing with mixing.
#[derive(Default)]
struct CharHasher(u64);

impl Hasher for CharHasher {
    #[inline(always)]
    fn finish(&self) -> u64 {
        self.0
    }

    #[inline(always)]
    fn write(&mut self, bytes: &[u8]) {
        // Fast path for 4-byte (char) writes - avoid loop
        if bytes.len() == 4 {
            let val = u32::from_ne_bytes([bytes[0], bytes[1], bytes[2], bytes[3]]);
            // Mix with golden ratio for better distribution
            self.0 = (val as u64).wrapping_mul(0x9E3779B97F4A7C15);
            return;
        }
        // Fallback for other sizes (shouldn't happen for char keys)
        for &byte in bytes {
            self.0 = self.0.wrapping_mul(0x100000001b3).wrapping_add(byte as u64);
        }
    }

    #[inline(always)]
    fn write_u32(&mut self, i: u32) {
        // Chars are u32 - use golden ratio multiplication for distribution
        self.0 = (i as u64).wrapping_mul(0x9E3779B97F4A7C15);
    }
}

type CharHasherBuilder = BuildHasherDefault<CharHasher>;

/// Advance a string slice by N characters, returning the remaining slice.
/// Returns empty string if n >= char count.
/// Optimized: uses direct byte slicing for ASCII, falls back to char_indices for UTF-8.
#[inline]
fn advance_by_chars(s: &str, n: usize) -> &str {
    if n == 0 {
        return s;
    }
    if n >= s.len() {
        return "";
    }
    // Fast path: if first N bytes are all ASCII, we can slice directly
    let bytes = s.as_bytes();
    if bytes[..n].is_ascii() {
        // Safe: we verified all bytes in [0..n] are ASCII (valid UTF-8 boundary)
        return &s[n..];
    }
    // Slow path: UTF-8 requires char-by-char traversal
    s.char_indices()
        .nth(n)
        .map(|(idx, _)| &s[idx..])
        .unwrap_or("")
}

/// Get the first N characters of a string as a new String.
/// More efficient than chars().take(n).collect() for known bounds.
#[inline]
fn take_chars(s: &str, n: usize) -> String {
    if n == 0 {
        return String::new();
    }
    s.char_indices()
        .nth(n)
        .map(|(idx, _)| s[..idx].to_string())
        .unwrap_or_else(|| s.to_string())
}

/// Node text with cached character count to avoid repeated O(n) chars().count() calls.
#[derive(Debug)]
struct NodeText {
    /// The actual text stored in this node
    text: String,
    /// Cached character count (UTF-8 chars, not bytes)
    char_count: usize,
}

impl NodeText {
    #[inline]
    fn new(text: String) -> Self {
        let char_count = text.chars().count();
        Self { text, char_count }
    }

    #[inline]
    fn empty() -> Self {
        Self {
            text: String::new(),
            char_count: 0,
        }
    }

    #[inline]
    fn char_count(&self) -> usize {
        self.char_count
    }

    #[inline]
    fn as_str(&self) -> &str {
        &self.text
    }

    #[inline]
    fn first_char(&self) -> Option<char> {
        self.text.chars().next()
    }

    /// Split the text at a character boundary, returning the prefix and suffix.
    /// This is more efficient than slice_by_chars as it computes both at once.
    #[inline]
    fn split_at_char(&self, char_idx: usize) -> (NodeText, NodeText) {
        if char_idx == 0 {
            return (NodeText::empty(), self.clone_text());
        }
        if char_idx >= self.char_count {
            return (self.clone_text(), NodeText::empty());
        }

        // Find byte index for the character boundary
        let byte_idx = self
            .text
            .char_indices()
            .nth(char_idx)
            .map(|(i, _)| i)
            .unwrap_or(self.text.len());

        let prefix = NodeText {
            text: self.text[..byte_idx].to_string(),
            char_count: char_idx,
        };
        let suffix = NodeText {
            text: self.text[byte_idx..].to_string(),
            char_count: self.char_count - char_idx,
        };
        (prefix, suffix)
    }

    #[inline]
    fn clone_text(&self) -> NodeText {
        NodeText {
            text: self.text.clone(),
            char_count: self.char_count,
        }
    }
}

impl Clone for NodeText {
    fn clone(&self) -> Self {
        self.clone_text()
    }
}

/// Global tenant string intern pool to avoid repeated allocations.
/// Uses DashMap for concurrent access with minimal contention.
static TENANT_INTERN_POOL: Lazy<DashMap<Arc<str>, ()>> = Lazy::new(DashMap::new);

/// Global epoch counter for LRU ordering.
/// Uses a simple incrementing counter instead of wall clock time.
///
/// Benefits:
/// - No syscall overhead (vs SystemTime::now())
/// - Smaller memory footprint (u64 vs u128)
/// - Perfectly monotonic (no clock skew issues)
///
/// For LRU eviction, relative ordering is all that matters.
static EPOCH_COUNTER: AtomicU64 = AtomicU64::new(0);

/// Get the next epoch value for LRU timestamp ordering.
/// Uses fetch_add for lock-free, monotonically increasing values.
/// Relaxed ordering is sufficient since we only need eventual consistency
/// for approximate LRU behavior.
#[inline]
fn get_epoch() -> u64 {
    EPOCH_COUNTER.fetch_add(1, Ordering::Relaxed)
}

#[derive(Debug)]
struct Node {
    /// Children nodes indexed by first character.
    /// Using custom hasher optimized for char keys.
    children: DashMap<char, NodeRef, CharHasherBuilder>,
    /// Node text with cached character count
    text: RwLock<NodeText>,
    /// Per-tenant last access epoch for LRU ordering. Using TenantId (Arc<str>) for cheap cloning.
    tenant_last_access_time: DashMap<TenantId, u64>,
    /// Parent pointer for upward traversal during timestamp updates.
    /// Uses Weak to avoid Arc reference cycles (parent -> child -> parent).
    parent: RwLock<Option<Weak<Node>>>,
    /// Cached last-accessed tenant for O(1) lookup during prefix match.
    /// Avoids O(shards) DashMap iteration in the common case.
    last_tenant: RwLock<Option<TenantId>>,
}

#[derive(Debug)]
pub struct Tree {
    root: NodeRef,
    /// Per-tenant character count for size tracking. Using TenantId for consistency.
    pub tenant_char_count: DashMap<TenantId, usize>,
}

// For the heap

struct EvictionEntry {
    timestamp: u64,
    tenant: TenantId,
    node: NodeRef,
}

impl Eq for EvictionEntry {}

#[allow(clippy::non_canonical_partial_ord_impl)]
impl PartialOrd for EvictionEntry {
    fn partial_cmp(&self, other: &Self) -> Option<std::cmp::Ordering> {
        Some(self.timestamp.cmp(&other.timestamp))
    }
}

impl Ord for EvictionEntry {
    fn cmp(&self, other: &Self) -> std::cmp::Ordering {
        self.timestamp.cmp(&other.timestamp)
    }
}

impl PartialEq for EvictionEntry {
    fn eq(&self, other: &Self) -> bool {
        self.timestamp == other.timestamp
    }
}

// For char operations
// Note that in rust, `.len()` or slice is operated on the "byte" level. It causes issues for UTF-8 characters because one character might use multiple bytes.
// https://en.wikipedia.org/wiki/UTF-8

/// Count matching prefix characters between two strings.
/// Returns the number of characters that match from the start.
/// Optimized: uses fast byte comparison for ASCII, falls back to char iteration for UTF-8.
#[inline]
fn shared_prefix_count(a: &str, b: &str) -> usize {
    let a_bytes = a.as_bytes();
    let b_bytes = b.as_bytes();

    // Find common byte prefix length using iterator (potentially SIMD-optimized)
    let common_byte_len = a_bytes
        .iter()
        .zip(b_bytes)
        .position(|(&a_byte, &b_byte)| a_byte != b_byte)
        .unwrap_or_else(|| a_bytes.len().min(b_bytes.len()));

    // If the common byte prefix is all ASCII, byte count == char count
    // Otherwise, fall back to char-by-char comparison for UTF-8 safety
    if a_bytes[..common_byte_len].is_ascii() {
        common_byte_len
    } else {
        shared_prefix_count_chars(a, b)
    }
}

/// Fallback char-by-char comparison for strings with non-ASCII characters.
#[inline]
fn shared_prefix_count_chars(a: &str, b: &str) -> usize {
    a.chars()
        .zip(b.chars())
        .take_while(|(a_char, b_char)| a_char == b_char)
        .count()
}

/// Intern tenant ID to avoid repeated allocations.
/// Returns cached Arc<str> if tenant was seen before.
#[inline]
fn intern_tenant(tenant: &str) -> TenantId {
    // Fast path: check if already interned
    if let Some(entry) = TENANT_INTERN_POOL.get(tenant) {
        return Arc::clone(entry.key());
    }

    // Slow path: intern new tenant
    let interned: Arc<str> = Arc::from(tenant);
    TENANT_INTERN_POOL.insert(Arc::clone(&interned), ());
    interned
}

impl Default for Tree {
    fn default() -> Self {
        Self::new()
    }
}

impl Tree {
    /*
    Thread-safe multi tenant radix tree

    1. Storing data for multiple tenants (the overlap of multiple radix tree)
    2. Node-level lock to enable concurrent access on nodes
    3. Leaf LRU eviction based on tenant access time

    Optimizations:
    - Cached character counts in NodeText to avoid O(n) chars().count() calls
    - Interned tenant IDs (Arc<str>) for cheap cloning and comparison
    - Batched timestamp updates to reduce syscalls
    - Custom hasher for char keys in children DashMap
    */

    pub fn new() -> Self {
        Tree {
            // Root uses higher shard count since ALL requests pass through it
            root: Arc::new(Node {
                children: DashMap::with_hasher_and_shard_amount(
                    CharHasherBuilder::default(),
                    ROOT_SHARD_COUNT,
                ),
                text: RwLock::new(NodeText::empty()),
                tenant_last_access_time: DashMap::with_shard_amount(ROOT_SHARD_COUNT),
                parent: RwLock::new(None),
                last_tenant: RwLock::new(None),
            }),
            tenant_char_count: DashMap::with_shard_amount(ROOT_SHARD_COUNT),
        }
    }

    pub fn insert_text(&self, text: &str, tenant: &str) {
        // Insert text into tree with given tenant
        // Use slice-based traversal to avoid Vec<char> allocation

        // Intern the tenant ID once for reuse
        let tenant_id = intern_tenant(tenant);

        // Ensure tenant exists at root (don't update timestamp - root is never evicted)
        self.root
            .tenant_last_access_time
            .entry(Arc::clone(&tenant_id))
            .or_insert(0);

        self.tenant_char_count
            .entry(Arc::clone(&tenant_id))
            .or_insert(0);

        // Track remaining text as a slice - no allocation needed
        let mut remaining = text;
        let mut prev = Arc::clone(&self.root);

        // Result type to carry state out of the match block
        // This allows the entry guard to be dropped before we update prev
        enum InsertStep {
            Done,
            Continue {
                next_prev: NodeRef,
                advance_chars: usize,
            },
        }

        while !remaining.is_empty() {
            let first_char = remaining.chars().next().unwrap();

            // Use entry API for atomic check-and-insert semantics (required for thread safety)
            let step = match prev.children.entry(first_char) {
                Entry::Vacant(entry) => {
                    // No match - create new node with remaining text (this is the leaf)
                    // Compute remaining char count lazily - only here when creating leaf
                    let remaining_char_count = remaining.chars().count();
                    let epoch = get_epoch();

                    let new_node = Arc::new(Node {
                        children: new_children_map(),
                        text: RwLock::new(NodeText::new(remaining.to_string())),
                        tenant_last_access_time: new_tenant_map(),
                        parent: RwLock::new(Some(Arc::downgrade(&prev))),
                        last_tenant: RwLock::new(Some(Arc::clone(&tenant_id))),
                    });

                    // Attach tenant to the new leaf node with timestamp
                    self.tenant_char_count
                        .entry(Arc::clone(&tenant_id))
                        .and_modify(|count| *count += remaining_char_count)
                        .or_insert(remaining_char_count);
                    new_node
                        .tenant_last_access_time
                        .insert(Arc::clone(&tenant_id), epoch);

                    entry.insert(new_node);
                    InsertStep::Done
                }

                Entry::Occupied(mut entry) => {
                    let matched_node = entry.get().clone();

                    let matched_node_text = matched_node.text.read();
                    let matched_node_text_count = matched_node_text.char_count();
                    let matched_node_text_str = matched_node_text.as_str();

                    // Use slice-based comparison - no allocation
                    let shared_count = shared_prefix_count(remaining, matched_node_text_str);

                    if shared_count < matched_node_text_count {
                        // Split the matched node
                        let (matched_text, contracted_text) =
                            matched_node_text.split_at_char(shared_count);
                        let matched_text_count = shared_count;

                        // Drop read lock before creating new node
                        drop(matched_node_text);

                        let new_node = Arc::new(Node {
                            text: RwLock::new(matched_text),
                            children: new_children_map(),
                            parent: RwLock::new(Some(Arc::downgrade(&prev))),
                            tenant_last_access_time: matched_node.tenant_last_access_time.clone(),
                            last_tenant: RwLock::new(matched_node.last_tenant.read().clone()),
                        });

                        let first_new_char = contracted_text.first_char().unwrap();
                        new_node
                            .children
                            .insert(first_new_char, Arc::clone(&matched_node));

                        entry.insert(Arc::clone(&new_node));

                        *matched_node.text.write() = contracted_text;
                        *matched_node.parent.write() = Some(Arc::downgrade(&new_node));

                        // Attach tenant to the new split node (intermediate - no timestamp update)
                        // The cloned DashMap already has the tenant; just ensure char count is correct
                        if !new_node
                            .tenant_last_access_time
                            .contains_key(tenant_id.as_ref())
                        {
                            self.tenant_char_count
                                .entry(Arc::clone(&tenant_id))
                                .and_modify(|count| *count += matched_text_count)
                                .or_insert(matched_text_count);
                            new_node
                                .tenant_last_access_time
                                .insert(Arc::clone(&tenant_id), 0);
                        }

                        InsertStep::Continue {
                            next_prev: new_node,
                            advance_chars: shared_count,
                        }
                    } else {
                        // Full match - move to next node (intermediate - no timestamp update)
                        drop(matched_node_text);

                        // Ensure tenant exists at this intermediate node
                        if !matched_node
                            .tenant_last_access_time
                            .contains_key(tenant_id.as_ref())
                        {
                            self.tenant_char_count
                                .entry(Arc::clone(&tenant_id))
                                .and_modify(|count| *count += matched_node_text_count)
                                .or_insert(matched_node_text_count);
                            matched_node
                                .tenant_last_access_time
                                .insert(Arc::clone(&tenant_id), 0);
                        }

                        InsertStep::Continue {
                            next_prev: matched_node,
                            advance_chars: shared_count,
                        }
                    }
                }
            };

            // Entry guard is now dropped - safe to update prev
            match step {
                InsertStep::Done => return, // New leaf created with timestamp, we're done
                InsertStep::Continue {
                    next_prev,
                    advance_chars,
                } => {
                    prev = next_prev;
                    remaining = advance_by_chars(remaining, advance_chars);
                }
            }
        }

        // Loop exited normally (remaining empty) - prev is the leaf node
        // Update its timestamp for LRU ordering
        let epoch = get_epoch();
        prev.tenant_last_access_time
            .insert(Arc::clone(&tenant_id), epoch);
    }

    /// Performs prefix matching and returns detailed result with char counts.
    /// Optimized: no string allocations, deferred char counting.
    pub fn match_prefix_with_counts(&self, text: &str) -> PrefixMatchResult {
        let mut remaining = text;
        let mut matched_chars = 0;
        let mut prev = Arc::clone(&self.root);

        while !remaining.is_empty() {
            let first_char = remaining.chars().next().unwrap();

            let child_node = prev.children.get(&first_char).map(|e| e.value().clone());

            if let Some(matched_node) = child_node {
                let matched_text_guard = matched_node.text.read();
                let matched_node_text_count = matched_text_guard.char_count();

                // Use slice-based comparison - no allocation
                let shared_count = shared_prefix_count(remaining, matched_text_guard.as_str());
                drop(matched_text_guard);

                if shared_count == matched_node_text_count {
                    // Full match with current node's text, continue to next node
                    matched_chars += shared_count;
                    remaining = advance_by_chars(remaining, shared_count);
                    prev = matched_node;
                } else {
                    // Partial match - still use this node for tenant selection
                    matched_chars += shared_count;
                    prev = matched_node;
                    break;
                }
            } else {
                // No match found, stop here
                break;
            }
        }

        let curr = prev;

        // Try cached tenant first (O(1)) before falling back to O(shards) DashMap iteration.
        // The cache is valid if the tenant still exists in tenant_last_access_time.
        let tenant: TenantId = {
            let cached = curr.last_tenant.read();
            if let Some(ref t) = *cached {
                if curr.tenant_last_access_time.contains_key(t.as_ref()) {
                    Arc::clone(t)
                } else {
                    drop(cached);
                    // Cache stale, fall back to iteration and update cache
                    let t = curr
                        .tenant_last_access_time
                        .iter()
                        .next()
                        .map(|kv| Arc::clone(kv.key()))
                        .unwrap_or_else(|| Arc::from("empty"));
                    *curr.last_tenant.write() = Some(Arc::clone(&t));
                    t
                }
            } else {
                drop(cached);
                // No cache, iterate and populate cache
                let t = curr
                    .tenant_last_access_time
                    .iter()
                    .next()
                    .map(|kv| Arc::clone(kv.key()))
                    .unwrap_or_else(|| Arc::from("empty"));
                *curr.last_tenant.write() = Some(Arc::clone(&t));
                t
            }
        };

        // Update timestamp probabilistically (1 in 8 matches) to reduce DashMap contention.
        // LRU eviction doesn't need perfect accuracy - approximate timestamps suffice.
        let epoch = get_epoch();
        if epoch & 0x7 == 0 {
            curr.tenant_last_access_time
                .insert(Arc::clone(&tenant), epoch);
        }

        // Compute input char count directly from input text.
        // This is equivalent to matched_chars + remaining.chars().count() but avoids
        // needing to track remaining precisely through the traversal.
        let input_char_count = text.chars().count();

        PrefixMatchResult {
            tenant,
            matched_char_count: matched_chars,
            input_char_count,
        }
    }

    /// Legacy prefix_match API for backward compatibility.
    /// Note: This computes matched_text which has allocation overhead.
    pub fn prefix_match_legacy(&self, text: &str) -> (String, String) {
        let result = self.match_prefix_with_counts(text);
        let matched_text = take_chars(text, result.matched_char_count);
        (matched_text, result.tenant.to_string())
    }

    #[allow(dead_code)]
    pub fn prefix_match_tenant(&self, text: &str, tenant: &str) -> String {
        // Use slice-based traversal - no Vec<char> allocation

        // Intern tenant ID once for efficient lookups
        let tenant_id = intern_tenant(tenant);

        let mut remaining = text;
        let mut matched_chars = 0;
        let mut prev = Arc::clone(&self.root);

        while !remaining.is_empty() {
            let first_char = remaining.chars().next().unwrap();

            let child_node = prev.children.get(&first_char).map(|e| e.value().clone());

            if let Some(matched_node) = child_node {
                // Only continue matching if this node belongs to the specified tenant
                if !matched_node
                    .tenant_last_access_time
                    .contains_key(tenant_id.as_ref())
                {
                    break;
                }

                let matched_text_guard = matched_node.text.read();
                let matched_node_text_count = matched_text_guard.char_count();

                // Use slice-based comparison - no allocation
                let shared_count = shared_prefix_count(remaining, matched_text_guard.as_str());
                drop(matched_text_guard);

                if shared_count == matched_node_text_count {
                    // Full match with current node's text, continue to next node
                    matched_chars += shared_count;
                    remaining = advance_by_chars(remaining, shared_count);
                    prev = matched_node;
                } else {
                    // Partial match - still use this node for timestamp update
                    matched_chars += shared_count;
                    prev = matched_node;
                    break;
                }
            } else {
                // No match found, stop here
                break;
            }
        }

        let curr = prev;

        // Only update timestamp if we found a match for the specified tenant.
        // Update matched node only - ancestor propagation is unnecessary.
        if curr
            .tenant_last_access_time
            .contains_key(tenant_id.as_ref())
        {
            let epoch = get_epoch();
            curr.tenant_last_access_time
                .insert(Arc::clone(&tenant_id), epoch);
        }

        // Build result from original input using char count
        take_chars(text, matched_chars)
    }

    /// Return the list of tenants for which this node is a leaf.
    /// A tenant is a leaf at this node if no children have that tenant.
    fn leaf_of(node: &NodeRef) -> Vec<TenantId> {
        let mut candidates: HashMap<TenantId, bool> = node
            .tenant_last_access_time
            .iter()
            .map(|entry| (Arc::clone(entry.key()), true))
            .collect();

        for child in node.children.iter() {
            for tenant in child.value().tenant_last_access_time.iter() {
                // Mark as non-leaf if any child has this tenant
                candidates.insert(Arc::clone(tenant.key()), false);
            }
        }

        candidates
            .into_iter()
            .filter(|(_, is_leaf)| *is_leaf)
            .map(|(tenant, _)| tenant)
            .collect()
    }

    pub fn evict_tenant_by_size(&self, max_size: usize) {
        // Calculate used size and collect leaves
        let mut stack = vec![Arc::clone(&self.root)];
        let mut pq = BinaryHeap::new();

        while let Some(curr) = stack.pop() {
            for child in curr.children.iter() {
                stack.push(Arc::clone(child.value()));
            }

            // Add leaves to priority queue
            for tenant in Tree::leaf_of(&curr) {
                if let Some(timestamp) = curr.tenant_last_access_time.get(tenant.as_ref()) {
                    pq.push(Reverse(EvictionEntry {
                        timestamp: *timestamp,
                        tenant: Arc::clone(&tenant),
                        node: Arc::clone(&curr),
                    }));
                }
            }
        }

        debug!("Before eviction - Used size per tenant:");
        for entry in self.tenant_char_count.iter() {
            debug!("Tenant: {}, Size: {}", entry.key(), entry.value());
        }

        // Process eviction
        while let Some(Reverse(entry)) = pq.pop() {
            let EvictionEntry { tenant, node, .. } = entry;

            if let Some(used_size) = self.tenant_char_count.get(tenant.as_ref()) {
                if *used_size <= max_size {
                    continue;
                }
            }

            // Verify this node is still a leaf for this tenant (may have changed)
            // A node is a leaf for a tenant if no children have that tenant
            let is_still_leaf = node.tenant_last_access_time.contains_key(tenant.as_ref())
                && !node.children.iter().any(|child| {
                    child
                        .value()
                        .tenant_last_access_time
                        .contains_key(tenant.as_ref())
                });
            if !is_still_leaf {
                continue;
            }

            // Decrement when removing tenant from node
            let node_len = node.text.read().char_count();
            self.tenant_char_count
                .entry(Arc::clone(&tenant))
                .and_modify(|count| {
                    *count = count.saturating_sub(node_len);
                });

            // Remove tenant from node
            node.tenant_last_access_time.remove(tenant.as_ref());

            // Get parent reference outside of the borrow scope
            // Use Weak::upgrade() to get a strong reference if the parent still exists
            let parent_opt = node.parent.read().as_ref().and_then(Weak::upgrade);

            // Remove empty nodes
            if node.children.is_empty() && node.tenant_last_access_time.is_empty() {
                if let Some(ref parent) = parent_opt {
                    if let Some(fc) = node.text.read().first_char() {
                        parent.children.remove(&fc);
                    }
                }
            }

            // If parent has this tenant and no other children have it,
            // parent becomes a new leaf - add to priority queue
            if let Some(ref parent) = parent_opt {
                if parent.tenant_last_access_time.contains_key(tenant.as_ref()) {
                    let has_child_with_tenant = parent.children.iter().any(|child| {
                        child
                            .value()
                            .tenant_last_access_time
                            .contains_key(tenant.as_ref())
                    });

                    if !has_child_with_tenant {
                        // Add parent to priority queue as new leaf
                        if let Some(timestamp) = parent.tenant_last_access_time.get(tenant.as_ref())
                        {
                            pq.push(Reverse(EvictionEntry {
                                timestamp: *timestamp,
                                tenant: Arc::clone(&tenant),
                                node: Arc::clone(parent),
                            }));
                        }
                    }
                }
            }
        }

        debug!("After eviction - Used size per tenant:");
        for entry in self.tenant_char_count.iter() {
            debug!("Tenant: {}, Size: {}", entry.key(), entry.value());
        }
    }

    // TODO: Implement efficient remove_tenant with reverse index.
    // See lib.rs for design options. Current naive O(n) traversal removed.
    // For now, stale entries are cleaned up by LRU eviction.

    #[allow(dead_code)]
    pub fn get_tenant_char_count(&self) -> HashMap<String, usize> {
        self.tenant_char_count
            .iter()
            .map(|entry| (entry.key().to_string(), *entry.value()))
            .collect()
    }

    #[allow(dead_code)]
    pub fn get_used_size_per_tenant(&self) -> HashMap<String, usize> {
        // perform a DFS to traverse all nodes and calculate the total size used by each tenant

        let mut used_size_per_tenant: HashMap<String, usize> = HashMap::new();
        let mut stack = vec![Arc::clone(&self.root)];

        while let Some(curr) = stack.pop() {
            // Use cached char count instead of chars().count()
            let text_count = curr.text.read().char_count();

            for tenant in curr.tenant_last_access_time.iter() {
                let size = used_size_per_tenant
                    .entry(tenant.key().to_string())
                    .or_insert(0);
                *size += text_count;
            }

            for child in curr.children.iter() {
                stack.push(Arc::clone(child.value()));
            }
        }

        used_size_per_tenant
    }

    /// Evict entries for a specific tenant to reduce to max_chars.
    pub fn evict_by_tenant(&self, tenant: &TenantId, max_chars: usize) {
        let current_count = self.tenant_char_count.get(tenant).map(|v| *v).unwrap_or(0);

        if current_count <= max_chars {
            return;
        }

        // Use existing eviction logic by temporarily setting a target
        // We need to evict (current_count - max_chars) chars
        self.evict_tenant_entries(tenant, current_count - max_chars);
    }

    /// Helper to evict a specific number of chars for a tenant
    fn evict_tenant_entries(&self, tenant_id: &TenantId, chars_to_evict: usize) {
        if chars_to_evict == 0 {
            return;
        }

        let mut evicted = 0;

        // Collect nodes with timestamps for LRU eviction
        let mut nodes_with_time: Vec<(NodeRef, u64)> = Vec::new();
        self.collect_tenant_nodes(&self.root, tenant_id, &mut nodes_with_time);

        // Sort by timestamp (oldest first)
        nodes_with_time.sort_by_key(|(_, ts)| *ts);

        for (node, _) in nodes_with_time {
            if evicted >= chars_to_evict {
                break;
            }

            let node_chars = node.text.read().char_count();
            if self.remove_tenant_from_node(&node, tenant_id) {
                evicted += node_chars;
            }
        }

        // Update tenant char count
        self.tenant_char_count
            .entry(tenant_id.clone())
            .and_modify(|count| *count = count.saturating_sub(evicted));
    }

    fn collect_tenant_nodes(
        &self,
        node: &NodeRef,
        tenant_id: &TenantId,
        result: &mut Vec<(NodeRef, u64)>,
    ) {
        // Skip root node as it should never be evicted
        if !Arc::ptr_eq(node, &self.root) {
            if let Some(ts) = node.tenant_last_access_time.get(tenant_id) {
                result.push((Arc::clone(node), *ts));
            }
        }

        for child in node.children.iter() {
            self.collect_tenant_nodes(child.value(), tenant_id, result);
        }
    }

    fn remove_tenant_from_node(&self, node: &NodeRef, tenant_id: &TenantId) -> bool {
        node.tenant_last_access_time.remove(tenant_id).is_some()
    }

    /// Get the char count for a specific tenant.
    pub fn tenant_char_size(&self, tenant: &TenantId) -> usize {
        self.tenant_char_count.get(tenant).map(|v| *v).unwrap_or(0)
    }

    /// Clear the tree to empty state.
    pub fn clear(&self) {
        // Clear root's children
        self.root.children.clear();
        // Clear root's tenant timestamps (except keep structure)
        self.root.tenant_last_access_time.clear();
        // Clear tenant char counts
        self.tenant_char_count.clear();
        // Reset root text
        *self.root.text.write() = NodeText::new(String::new());
    }

    #[allow(dead_code)]
    fn node_to_string(node: &NodeRef, prefix: &str, is_last: bool) -> String {
        let mut result = String::new();

        // Add prefix and branch character
        result.push_str(prefix);
        result.push_str(if is_last { "└── " } else { "├── " });

        // Add node text
        let node_text = node.text.read();
        result.push_str(&format!("'{}' [", node_text.as_str()));

        // Add tenant information with epoch values
        let mut tenant_info = Vec::new();
        for entry in node.tenant_last_access_time.iter() {
            let tenant_id = entry.key();
            let epoch = entry.value();
            tenant_info.push(format!("{} | epoch:{}", tenant_id, epoch));
        }

        result.push_str(&tenant_info.join(", "));
        result.push_str("]\n");

        // Process children
        let children: Vec<_> = node.children.iter().collect();
        let child_count = children.len();

        for (i, entry) in children.iter().enumerate() {
            let is_last_child = i == child_count - 1;
            let new_prefix = format!("{}{}", prefix, if is_last { "    " } else { "│   " });

            result.push_str(&Tree::node_to_string(
                entry.value(),
                &new_prefix,
                is_last_child,
            ));
        }

        result
    }

    #[allow(dead_code)]
    pub fn pretty_print(&self) {
        if self.root.children.is_empty() {
            return;
        }

        let mut result = String::new();
        let children: Vec<_> = self.root.children.iter().collect();
        let child_count = children.len();

        for (i, entry) in children.iter().enumerate() {
            let is_last = i == child_count - 1;
            result.push_str(&Tree::node_to_string(entry.value(), "", is_last));
        }

        println!("{result}");
    }
}

impl RadixTree for Tree {
    type Key = str;
    type MatchResult = PrefixMatchResult;

    fn insert(&self, key: &Self::Key, tenant: &str) {
        self.insert_text(key, tenant);
    }

    fn prefix_match(&self, key: &Self::Key) -> Option<TenantId> {
        let result = self.match_prefix_with_counts(key);
        if result.matched_char_count > 0 {
            Some(result.tenant)
        } else {
            None
        }
    }

    fn prefix_match_with_counts(&self, key: &Self::Key) -> Self::MatchResult {
        self.match_prefix_with_counts(key)
    }

    fn evict(&self, tenant: &TenantId, max_units: usize) {
        self.evict_by_tenant(tenant, max_units);
    }

    fn tenant_size(&self, tenant: &TenantId) -> usize {
        self.tenant_char_size(tenant)
    }

    fn reset(&self) {
        self.clear();
    }
}

//  Unit tests
#[cfg(test)]
mod tests {
    use std::{
        thread,
        time::{Duration, Instant},
    };

    use rand::{
        distr::{Alphanumeric, SampleString},
        rng as thread_rng, Rng,
    };

    use super::*;

    /// Helper to convert tenant_char_count to HashMap<String, usize> for comparison
    fn get_maintained_counts(tree: &Tree) -> HashMap<String, usize> {
        tree.tenant_char_count
            .iter()
            .map(|entry| (entry.key().to_string(), *entry.value()))
            .collect()
    }

    #[test]
    fn test_tenant_char_count() {
        let tree = Tree::new();

        tree.insert_text("apple", "tenant1");
        tree.insert_text("apricot", "tenant1");
        tree.insert_text("banana", "tenant1");
        tree.insert_text("amplify", "tenant2");
        tree.insert_text("application", "tenant2");

        let computed_sizes = tree.get_used_size_per_tenant();
        let maintained_counts = get_maintained_counts(&tree);

        println!("Phase 1 - Maintained vs Computed counts:");
        println!(
            "Maintained: {:?}\nComputed: {:?}",
            maintained_counts, computed_sizes
        );
        assert_eq!(
            maintained_counts, computed_sizes,
            "Phase 1: Initial insertions"
        );

        tree.insert_text("apartment", "tenant1");
        tree.insert_text("appetite", "tenant2");
        tree.insert_text("ball", "tenant1");
        tree.insert_text("box", "tenant2");

        let computed_sizes = tree.get_used_size_per_tenant();
        let maintained_counts = get_maintained_counts(&tree);

        println!("Phase 2 - Maintained vs Computed counts:");
        println!(
            "Maintained: {:?}\nComputed: {:?}",
            maintained_counts, computed_sizes
        );
        assert_eq!(
            maintained_counts, computed_sizes,
            "Phase 2: Additional insertions"
        );

        tree.insert_text("zebra", "tenant1");
        tree.insert_text("zebra", "tenant2");
        tree.insert_text("zero", "tenant1");
        tree.insert_text("zero", "tenant2");

        let computed_sizes = tree.get_used_size_per_tenant();
        let maintained_counts = get_maintained_counts(&tree);

        println!("Phase 3 - Maintained vs Computed counts:");
        println!(
            "Maintained: {:?}\nComputed: {:?}",
            maintained_counts, computed_sizes
        );
        assert_eq!(
            maintained_counts, computed_sizes,
            "Phase 3: Overlapping insertions"
        );

        tree.evict_tenant_by_size(10);

        let computed_sizes = tree.get_used_size_per_tenant();
        let maintained_counts = get_maintained_counts(&tree);

        println!("Phase 4 - Maintained vs Computed counts:");
        println!(
            "Maintained: {:?}\nComputed: {:?}",
            maintained_counts, computed_sizes
        );
        assert_eq!(maintained_counts, computed_sizes, "Phase 4: After eviction");
    }

    fn random_string(len: usize) -> String {
        Alphanumeric.sample_string(&mut thread_rng(), len)
    }

    #[test]
    fn test_cold_start() {
        let tree = Tree::new();

        let (matched_text, tenant) = tree.prefix_match_legacy("hello");

        assert_eq!(matched_text, "");
        assert_eq!(tenant, "empty");
    }

    #[test]
    fn test_exact_match_seq() {
        let tree = Tree::new();
        tree.insert_text("hello", "tenant1");
        tree.pretty_print();
        tree.insert_text("apple", "tenant2");
        tree.pretty_print();
        tree.insert_text("banana", "tenant3");
        tree.pretty_print();

        let (matched_text, tenant) = tree.prefix_match_legacy("hello");
        assert_eq!(matched_text, "hello");
        assert_eq!(tenant, "tenant1");

        let (matched_text, tenant) = tree.prefix_match_legacy("apple");
        assert_eq!(matched_text, "apple");
        assert_eq!(tenant, "tenant2");

        let (matched_text, tenant) = tree.prefix_match_legacy("banana");
        assert_eq!(matched_text, "banana");
        assert_eq!(tenant, "tenant3");
    }

    #[test]
    fn test_exact_match_concurrent() {
        let tree = Arc::new(Tree::new());

        // spawn 3 threads for insert
        let tree_clone = Arc::clone(&tree);

        let texts = ["hello", "apple", "banana"];
        let tenants = ["tenant1", "tenant2", "tenant3"];

        let mut handles = vec![];

        for i in 0..3 {
            let tree_clone = Arc::clone(&tree_clone);
            let text = texts[i];
            let tenant = tenants[i];

            let handle = thread::spawn(move || {
                tree_clone.insert_text(text, tenant);
            });

            handles.push(handle);
        }

        // wait
        for handle in handles {
            handle.join().unwrap();
        }

        // spawn 3 threads for match
        let mut handles = vec![];

        let tree_clone = Arc::clone(&tree);

        for i in 0..3 {
            let tree_clone = Arc::clone(&tree_clone);
            let text = texts[i];
            let tenant = tenants[i];

            let handle = thread::spawn(move || {
                let (matched_text, matched_tenant) = tree_clone.prefix_match_legacy(text);
                assert_eq!(matched_text, text);
                assert_eq!(matched_tenant, tenant);
            });

            handles.push(handle);
        }

        // wait
        for handle in handles {
            handle.join().unwrap();
        }
    }

    #[test]
    fn test_partial_match_concurrent() {
        let tree = Arc::new(Tree::new());

        // spawn 3 threads for insert
        let tree_clone = Arc::clone(&tree);

        static TEXTS: [&str; 3] = ["apple", "apabc", "acbdeds"];

        let mut handles = vec![];

        for text in TEXTS.iter() {
            let tree_clone = Arc::clone(&tree_clone);
            let tenant = "tenant0";

            let handle = thread::spawn(move || {
                tree_clone.insert_text(text, tenant);
            });

            handles.push(handle);
        }

        // wait
        for handle in handles {
            handle.join().unwrap();
        }

        // spawn 3 threads for match
        let mut handles = vec![];

        let tree_clone = Arc::clone(&tree);

        for text in TEXTS.iter() {
            let tree_clone = Arc::clone(&tree_clone);
            let tenant = "tenant0";

            let handle = thread::spawn(move || {
                let (matched_text, matched_tenant) = tree_clone.prefix_match_legacy(text);
                assert_eq!(matched_text, *text);
                assert_eq!(matched_tenant, tenant);
            });

            handles.push(handle);
        }

        // wait
        for handle in handles {
            handle.join().unwrap();
        }
    }

    #[test]
    fn test_group_prefix_insert_match_concurrent() {
        static PREFIXES: [&str; 4] = [
            "Clock strikes midnight, I'm still wide awake",
            "Got dreams bigger than these city lights",
            "Time waits for no one, gotta make my move",
            "Started from the bottom, that's no metaphor",
        ];
        let suffixes = [
            "Got too much to prove, ain't got time to lose",
            "History in the making, yeah, you can't erase this",
        ];
        let tree = Arc::new(Tree::new());

        let mut handles = vec![];

        for (i, prefix) in PREFIXES.iter().enumerate() {
            for suffix in suffixes.iter() {
                let tree_clone = Arc::clone(&tree);
                let text = format!("{} {}", prefix, suffix);
                let tenant = format!("tenant{}", i);

                let handle = thread::spawn(move || {
                    tree_clone.insert_text(&text, &tenant);
                });

                handles.push(handle);
            }
        }

        // wait
        for handle in handles {
            handle.join().unwrap();
        }

        tree.pretty_print();

        // check matching using multi threads
        let mut handles = vec![];

        for (i, prefix) in PREFIXES.iter().enumerate() {
            let tree_clone = Arc::clone(&tree);

            let handle = thread::spawn(move || {
                let (matched_text, matched_tenant) = tree_clone.prefix_match_legacy(prefix);
                let tenant = format!("tenant{}", i);
                assert_eq!(matched_text, *prefix);
                assert_eq!(matched_tenant, tenant);
            });

            handles.push(handle);
        }

        // wait
        for handle in handles {
            handle.join().unwrap();
        }
    }

    #[test]
    fn test_mixed_concurrent_insert_match() {
        // ensure it does not deadlock instead of doing correctness check

        static PREFIXES: [&str; 4] = [
            "Clock strikes midnight, I'm still wide awake",
            "Got dreams bigger than these city lights",
            "Time waits for no one, gotta make my move",
            "Started from the bottom, that's no metaphor",
        ];
        let suffixes = [
            "Got too much to prove, ain't got time to lose",
            "History in the making, yeah, you can't erase this",
        ];
        let tree = Arc::new(Tree::new());

        let mut handles = vec![];

        for (i, prefix) in PREFIXES.iter().enumerate() {
            for suffix in suffixes.iter() {
                let tree_clone = Arc::clone(&tree);
                let text = format!("{} {}", prefix, suffix);
                let tenant = format!("tenant{}", i);

                let handle = thread::spawn(move || {
                    tree_clone.insert_text(&text, &tenant);
                });

                handles.push(handle);
            }
        }

        // check matching using multi threads
        for prefix in PREFIXES.iter() {
            let tree_clone = Arc::clone(&tree);

            let handle = thread::spawn(move || {
                let (_matched_text, _matched_tenant) = tree_clone.prefix_match_legacy(prefix);
            });

            handles.push(handle);
        }

        // wait
        for handle in handles {
            handle.join().unwrap();
        }
    }

    #[test]
    fn test_utf8_split_seq() {
        // The string should be indexed and split by a utf-8 value basis instead of byte basis
        // use .chars() to get the iterator of the utf-8 value
        let tree = Arc::new(Tree::new());

        static TEST_PAIRS: [(&str, &str); 3] = [
            ("你好嗎", "tenant1"),
            ("你好喔", "tenant2"),
            ("你心情好嗎", "tenant3"),
        ];

        // Insert sequentially
        for (text, tenant) in TEST_PAIRS.iter() {
            tree.insert_text(text, tenant);
        }

        tree.pretty_print();

        for (text, tenant) in TEST_PAIRS.iter() {
            let (matched_text, matched_tenant) = tree.prefix_match_legacy(text);
            assert_eq!(matched_text, *text);
            assert_eq!(matched_tenant, *tenant);
        }
    }

    #[test]
    fn test_utf8_split_concurrent() {
        let tree = Arc::new(Tree::new());

        static TEST_PAIRS: [(&str, &str); 3] = [
            ("你好嗎", "tenant1"),
            ("你好喔", "tenant2"),
            ("你心情好嗎", "tenant3"),
        ];

        // Create multiple threads for insertion
        let mut handles = vec![];

        for (text, tenant) in TEST_PAIRS.iter() {
            let tree_clone = Arc::clone(&tree);

            let handle = thread::spawn(move || {
                tree_clone.insert_text(text, tenant);
            });

            handles.push(handle);
        }

        // Wait for all insertions to complete
        for handle in handles {
            handle.join().unwrap();
        }

        tree.pretty_print();

        // Create multiple threads for matching
        let mut handles = vec![];

        for (text, tenant) in TEST_PAIRS.iter() {
            let tree_clone = Arc::clone(&tree);

            let handle = thread::spawn(move || {
                let (matched_text, matched_tenant) = tree_clone.prefix_match_legacy(text);
                assert_eq!(matched_text, *text);
                assert_eq!(matched_tenant, *tenant);
            });

            handles.push(handle);
        }

        // Wait for all matches to complete
        for handle in handles {
            handle.join().unwrap();
        }
    }

    #[test]
    fn test_simple_eviction() {
        let tree = Tree::new();
        let max_size = 5;

        // Insert strings for both tenants
        tree.insert_text("hello", "tenant1"); // size 5

        tree.insert_text("hello", "tenant2"); // size 5
        thread::sleep(Duration::from_millis(10));
        tree.insert_text("world", "tenant2"); // size 5, total for tenant2 = 10

        tree.pretty_print();

        let sizes_before = tree.get_used_size_per_tenant();
        assert_eq!(sizes_before.get("tenant1").unwrap(), &5); // "hello" = 5
        assert_eq!(sizes_before.get("tenant2").unwrap(), &10); // "hello" + "world" = 10

        // Evict - should remove "hello" from tenant2 as it's the oldest
        tree.evict_tenant_by_size(max_size);

        tree.pretty_print();

        let sizes_after = tree.get_used_size_per_tenant();
        assert_eq!(sizes_after.get("tenant1").unwrap(), &5); // Should be unchanged
        assert_eq!(sizes_after.get("tenant2").unwrap(), &5); // Only "world" remains

        let (matched, tenant) = tree.prefix_match_legacy("world");
        assert_eq!(matched, "world");
        assert_eq!(tenant, "tenant2");
    }

    #[test]
    fn test_advanced_eviction() {
        let tree = Tree::new();

        // Set limits for each tenant
        let max_size: usize = 100;

        // Define prefixes
        let prefixes = ["aqwefcisdf", "iajsdfkmade", "kjnzxcvewqe", "iejksduqasd"];

        // Insert strings with shared prefixes
        for _i in 0..100 {
            for (j, prefix) in prefixes.iter().enumerate() {
                let random_suffix = random_string(10);
                let text = format!("{}{}", prefix, random_suffix);
                let tenant = format!("tenant{}", j + 1);
                tree.insert_text(&text, &tenant);
            }
        }

        // Perform eviction
        tree.evict_tenant_by_size(max_size);

        // Check sizes after eviction
        let sizes_after = tree.get_used_size_per_tenant();
        for (tenant, &size) in sizes_after.iter() {
            assert!(
                size <= max_size,
                "Tenant {} exceeds size limit. Current size: {}, Limit: {}",
                tenant,
                size,
                max_size
            );
        }
    }

    #[test]
    fn test_concurrent_operations_with_eviction() {
        // Ensure eviction works fine with concurrent insert and match operations for a given period

        let tree = Arc::new(Tree::new());
        let mut handles = vec![];
        let test_duration = Duration::from_secs(10);
        let start_time = Instant::now();
        let max_size = 100; // Single max size for all tenants

        // Spawn eviction thread
        {
            let tree = Arc::clone(&tree);
            let handle = thread::spawn(move || {
                while start_time.elapsed() < test_duration {
                    // Run eviction
                    tree.evict_tenant_by_size(max_size);

                    // Sleep for 5 seconds
                    thread::sleep(Duration::from_secs(5));
                }
            });
            handles.push(handle);
        }

        // Spawn 4 worker threads
        for thread_id in 0..4 {
            let tree = Arc::clone(&tree);
            let handle = thread::spawn(move || {
                let mut rng = rand::rng();
                let tenant = format!("tenant{}", thread_id + 1);
                let prefix = format!("prefix{}", thread_id);

                while start_time.elapsed() < test_duration {
                    // Random decision: match or insert (70% match, 30% insert)
                    if rng.random_bool(0.7) {
                        // Perform match operation
                        let random_len = rng.random_range(3..10);
                        let search_str = format!("{}{}", prefix, random_string(random_len));
                        let (_matched, _) = tree.prefix_match_legacy(&search_str);
                    } else {
                        // Perform insert operation
                        let random_len = rng.random_range(5..15);
                        let insert_str = format!("{}{}", prefix, random_string(random_len));
                        tree.insert_text(&insert_str, &tenant);
                        // println!("Thread {} inserted: {}", thread_id, insert_str);
                    }

                    // Small random sleep to vary timing
                    thread::sleep(Duration::from_millis(rng.random_range(10..100)));
                }
            });
            handles.push(handle);
        }

        // Wait for all threads to complete
        for handle in handles {
            handle.join().unwrap();
        }

        // final eviction
        tree.evict_tenant_by_size(max_size);

        // Final size check
        let final_sizes = tree.get_used_size_per_tenant();
        println!("Final sizes after test completion: {:?}", final_sizes);

        for (_, &size) in final_sizes.iter() {
            assert!(
                size <= max_size,
                "Tenant exceeds size limit. Final size: {}, Limit: {}",
                size,
                max_size
            );
        }
    }

    #[test]
    fn test_leaf_of() {
        let tree = Tree::new();

        // Helper to convert leaves to strings for easier assertion
        let leaves_as_strings =
            |leaves: &[TenantId]| -> Vec<String> { leaves.iter().map(|t| t.to_string()).collect() };

        // Single node
        tree.insert_text("hello", "tenant1");
        let leaves = Tree::leaf_of(&tree.root.children.get(&'h').unwrap());
        assert_eq!(leaves_as_strings(&leaves), vec!["tenant1"]);

        // Node with multiple tenants
        tree.insert_text("hello", "tenant2");
        let leaves = Tree::leaf_of(&tree.root.children.get(&'h').unwrap());
        let leaves_str = leaves_as_strings(&leaves);
        assert_eq!(leaves_str.len(), 2);
        assert!(leaves_str.contains(&"tenant1".to_string()));
        assert!(leaves_str.contains(&"tenant2".to_string()));

        // Non-leaf node
        tree.insert_text("hi", "tenant1");
        let leaves = Tree::leaf_of(&tree.root.children.get(&'h').unwrap());
        assert!(leaves.is_empty());
    }

    #[test]
    fn test_get_used_size_per_tenant() {
        let tree = Tree::new();

        // Single tenant
        tree.insert_text("hello", "tenant1");
        tree.insert_text("world", "tenant1");
        let sizes = tree.get_used_size_per_tenant();

        tree.pretty_print();
        println!("{:?}", sizes);
        assert_eq!(sizes.get("tenant1").unwrap(), &10); // "hello" + "world"

        // Multiple tenants sharing nodes
        tree.insert_text("hello", "tenant2");
        tree.insert_text("help", "tenant2");
        let sizes = tree.get_used_size_per_tenant();

        tree.pretty_print();
        println!("{:?}", sizes);
        assert_eq!(sizes.get("tenant1").unwrap(), &10);
        assert_eq!(sizes.get("tenant2").unwrap(), &6); // "hello" + "p"

        // UTF-8 characters
        tree.insert_text("你好", "tenant3");
        let sizes = tree.get_used_size_per_tenant();
        tree.pretty_print();
        println!("{:?}", sizes);
        assert_eq!(sizes.get("tenant3").unwrap(), &2); // 2 Chinese characters

        tree.pretty_print();
    }

    #[test]
    fn test_prefix_match_tenant() {
        let tree = Tree::new();

        // Insert overlapping prefixes for different tenants
        tree.insert_text("hello", "tenant1"); // tenant1: hello
        tree.insert_text("hello", "tenant2"); // tenant2: hello
        tree.insert_text("hello world", "tenant2"); // tenant2: hello -> world
        tree.insert_text("help", "tenant1"); // tenant1: hel -> p
        tree.insert_text("helicopter", "tenant2"); // tenant2: hel -> icopter

        assert_eq!(tree.prefix_match_tenant("hello", "tenant1"), "hello"); // Full match for tenant1
        assert_eq!(tree.prefix_match_tenant("help", "tenant1"), "help"); // Exclusive to tenant1
        assert_eq!(tree.prefix_match_tenant("hel", "tenant1"), "hel"); // Shared prefix
        assert_eq!(tree.prefix_match_tenant("hello world", "tenant1"), "hello"); // Should stop at tenant1's boundary
        assert_eq!(tree.prefix_match_tenant("helicopter", "tenant1"), "hel"); // Should stop at tenant1's boundary

        assert_eq!(tree.prefix_match_tenant("hello", "tenant2"), "hello"); // Full match for tenant2
        assert_eq!(
            tree.prefix_match_tenant("hello world", "tenant2"),
            "hello world"
        ); // Exclusive to tenant2
        assert_eq!(
            tree.prefix_match_tenant("helicopter", "tenant2"),
            "helicopter"
        ); // Exclusive to tenant2
        assert_eq!(tree.prefix_match_tenant("hel", "tenant2"), "hel"); // Shared prefix
        assert_eq!(tree.prefix_match_tenant("help", "tenant2"), "hel"); // Should stop at tenant2's boundary

        assert_eq!(tree.prefix_match_tenant("hello", "tenant3"), ""); // Non-existent tenant
        assert_eq!(tree.prefix_match_tenant("help", "tenant3"), ""); // Non-existent tenant
    }

    // NOTE: test_simple_tenant_eviction and test_complex_tenant_eviction removed
    // because remove_tenant was removed (inefficient O(n) implementation).
    // TODO: Re-add these tests when efficient remove_tenant is implemented.

    // ==================== Edge Case Tests ====================

    #[test]
    fn test_empty_string_input() {
        let tree = Tree::new();

        // Insert empty string
        tree.insert_text("", "tenant1");

        // Match empty string
        let (matched, tenant) = tree.prefix_match_legacy("");
        assert_eq!(matched, "");
        assert_eq!(tenant, "tenant1");

        // Insert non-empty, then match empty
        tree.insert_text("hello", "tenant2");
        let (matched, tenant) = tree.prefix_match_legacy("");
        assert_eq!(matched, "");
        assert_eq!(tenant, "tenant1");
    }

    #[test]
    fn test_single_character_operations() {
        let tree = Tree::new();

        // Insert single characters
        tree.insert_text("a", "tenant1");
        tree.insert_text("b", "tenant2");
        tree.insert_text("c", "tenant1");

        let (matched, tenant) = tree.prefix_match_legacy("a");
        assert_eq!(matched, "a");
        assert_eq!(tenant, "tenant1");

        let (matched, tenant) = tree.prefix_match_legacy("b");
        assert_eq!(matched, "b");
        assert_eq!(tenant, "tenant2");

        // Match with longer string starting with single char
        let (matched, tenant) = tree.prefix_match_legacy("abc");
        assert_eq!(matched, "a");
        assert_eq!(tenant, "tenant1");
    }

    #[test]
    fn test_prefix_is_subset_of_existing() {
        let tree = Tree::new();

        // Insert longer string first
        tree.insert_text("application", "tenant1");

        // Now insert prefix of existing
        tree.insert_text("app", "tenant2");

        // Match the prefix - both tenants own "app" node
        let (matched, tenant) = tree.prefix_match_legacy("app");
        assert_eq!(matched, "app");
        assert!(tenant == "tenant1" || tenant == "tenant2");

        // Match longer string
        let (matched, tenant) = tree.prefix_match_legacy("application");
        assert_eq!(matched, "application");
        assert_eq!(tenant, "tenant1");

        // Match "apple" - matches "app" + "l" from the child node = "appl"
        // Then 'e' doesn't match 'i' in the remaining suffix, so stops at 4 chars
        let (matched, _tenant) = tree.prefix_match_legacy("apple");
        assert_eq!(matched, "appl");
    }

    #[test]
    fn test_existing_is_prefix_of_new() {
        let tree = Tree::new();

        // Insert shorter string first
        tree.insert_text("app", "tenant1");

        // Now insert longer string with same prefix
        tree.insert_text("application", "tenant2");

        let (matched, tenant) = tree.prefix_match_legacy("app");
        assert_eq!(matched, "app");
        assert!(tenant == "tenant1" || tenant == "tenant2");

        let (matched, tenant) = tree.prefix_match_legacy("application");
        assert_eq!(matched, "application");
        assert_eq!(tenant, "tenant2");

        // "applesauce" matches "app" + "l" from the child node = "appl"
        // Then 'e' in "esauce" doesn't match 'i' in the suffix, so matching stops
        let (matched, _tenant) = tree.prefix_match_legacy("applesauce");
        assert_eq!(matched, "appl");
    }

    // ==================== prefix_match_with_counts Tests ====================

    #[test]
    fn test_prefix_match_with_counts_accuracy() {
        let tree = Tree::new();

        tree.insert_text("hello world", "tenant1");

        // Exact match
        let result = tree.match_prefix_with_counts("hello world");
        assert_eq!(result.matched_char_count, 11);
        assert_eq!(result.input_char_count, 11);
        assert_eq!(&*result.tenant, "tenant1");

        // Partial match
        let result = tree.match_prefix_with_counts("hello");
        assert_eq!(result.matched_char_count, 5);
        assert_eq!(result.input_char_count, 5);

        // Extended match
        let result = tree.match_prefix_with_counts("hello world and more");
        assert_eq!(result.matched_char_count, 11);
        assert_eq!(result.input_char_count, 20);

        // No match
        let result = tree.match_prefix_with_counts("goodbye");
        assert_eq!(result.matched_char_count, 0);
        assert_eq!(result.input_char_count, 7);
    }

    #[test]
    fn test_prefix_match_with_counts_utf8() {
        let tree = Tree::new();

        // UTF-8 string: 5 characters, more bytes
        tree.insert_text("你好世界呀", "tenant1");

        let result = tree.match_prefix_with_counts("你好世界呀");
        assert_eq!(result.matched_char_count, 5);
        assert_eq!(result.input_char_count, 5);

        let result = tree.match_prefix_with_counts("你好");
        assert_eq!(result.matched_char_count, 2);
        assert_eq!(result.input_char_count, 2);

        // Mixed ASCII and UTF-8
        tree.insert_text("hello你好", "tenant2");
        let result = tree.match_prefix_with_counts("hello你好世界");
        assert_eq!(result.matched_char_count, 7); // "hello你好" = 7 chars
        assert_eq!(result.input_char_count, 9); // "hello你好世界" = 9 chars
    }

    // ==================== Node Splitting Edge Cases ====================

    #[test]
    fn test_split_at_first_character() {
        let tree = Tree::new();

        // Insert "abc"
        tree.insert_text("abc", "tenant1");

        // Insert "aXX" - should split at first char
        tree.insert_text("aXX", "tenant2");

        let (matched, tenant) = tree.prefix_match_legacy("abc");
        assert_eq!(matched, "abc");
        assert_eq!(tenant, "tenant1");

        let (matched, tenant) = tree.prefix_match_legacy("aXX");
        assert_eq!(matched, "aXX");
        assert_eq!(tenant, "tenant2");

        let (matched, _) = tree.prefix_match_legacy("a");
        assert_eq!(matched, "a");
    }

    #[test]
    fn test_split_at_last_character() {
        let tree = Tree::new();

        // Insert "abcd"
        tree.insert_text("abcd", "tenant1");

        // Insert "abcX" - should split at last char of shared prefix
        tree.insert_text("abcX", "tenant2");

        let (matched, tenant) = tree.prefix_match_legacy("abcd");
        assert_eq!(matched, "abcd");
        assert_eq!(tenant, "tenant1");

        let (matched, tenant) = tree.prefix_match_legacy("abcX");
        assert_eq!(matched, "abcX");
        assert_eq!(tenant, "tenant2");

        let (matched, _) = tree.prefix_match_legacy("abc");
        assert_eq!(matched, "abc");
    }

    #[test]
    fn test_multiple_splits_same_path() {
        let tree = Tree::new();

        // Create a chain of splits
        tree.insert_text("abcdefgh", "tenant1");
        tree.insert_text("abcdef", "tenant2");
        tree.insert_text("abcd", "tenant3");
        tree.insert_text("ab", "tenant4");

        // Verify all paths work
        assert_eq!(tree.prefix_match_legacy("abcdefgh").0, "abcdefgh");
        assert_eq!(tree.prefix_match_legacy("abcdef").0, "abcdef");
        assert_eq!(tree.prefix_match_legacy("abcd").0, "abcd");
        assert_eq!(tree.prefix_match_legacy("ab").0, "ab");
        assert_eq!(tree.prefix_match_legacy("a").0, "a");
    }

    // ==================== High Contention Stress Tests ====================

    #[test]
    fn test_high_contention_same_prefix() {
        let tree = Arc::new(Tree::new());
        let num_threads = 16;
        let ops_per_thread = 100;
        let mut handles = vec![];

        // All threads operate on strings with same prefix
        for thread_id in 0..num_threads {
            let tree = Arc::clone(&tree);
            let handle = thread::spawn(move || {
                let tenant = format!("tenant{}", thread_id);
                for i in 0..ops_per_thread {
                    let text = format!("shared_prefix_{}", i);
                    tree.insert_text(&text, &tenant);

                    // Immediately try to match
                    let (matched, _) = tree.prefix_match_legacy(&text);
                    assert!(
                        matched.starts_with("shared_prefix_"),
                        "Match should start with shared_prefix_"
                    );
                }
            });
            handles.push(handle);
        }

        for handle in handles {
            handle.join().expect("Thread panicked");
        }

        // Verify tree is still consistent
        let sizes = tree.get_used_size_per_tenant();
        assert!(!sizes.is_empty(), "Tree should have entries");
    }

    // NOTE: test_rapid_insert_remove_cycles removed because remove_tenant was removed.
    // TODO: Re-add this test when efficient remove_tenant is implemented.

    // ==================== ASCII/UTF-8 Consistency Tests ====================

    #[test]
    fn test_ascii_utf8_consistency() {
        let tree = Tree::new();

        // Insert ASCII
        tree.insert_text("hello", "tenant1");

        // Insert UTF-8 with same logical prefix (none)
        tree.insert_text("你好", "tenant2");

        // Insert mixed
        tree.insert_text("hello你好", "tenant3");

        // All should be retrievable
        assert_eq!(tree.prefix_match_legacy("hello").0, "hello");
        assert_eq!(tree.prefix_match_legacy("你好").0, "你好");
        assert_eq!(tree.prefix_match_legacy("hello你好").0, "hello你好");

        // Counts should be correct
        let result = tree.match_prefix_with_counts("hello");
        assert_eq!(result.matched_char_count, 5);
        assert_eq!(result.input_char_count, 5);

        let result = tree.match_prefix_with_counts("你好");
        assert_eq!(result.matched_char_count, 2);
        assert_eq!(result.input_char_count, 2);

        let result = tree.match_prefix_with_counts("hello你好");
        assert_eq!(result.matched_char_count, 7);
        assert_eq!(result.input_char_count, 7);
    }

    #[test]
    fn test_emoji_handling() {
        let tree = Tree::new();

        // Emoji are multi-byte UTF-8
        tree.insert_text("hello 👋", "tenant1");
        tree.insert_text("hello 👋🌍", "tenant2");

        let (matched, tenant) = tree.prefix_match_legacy("hello 👋");
        assert_eq!(matched, "hello 👋");
        assert_eq!(tenant, "tenant1");

        let (matched, tenant) = tree.prefix_match_legacy("hello 👋🌍");
        assert_eq!(matched, "hello 👋🌍");
        assert_eq!(tenant, "tenant2");

        // Verify char count (not byte count)
        let result = tree.match_prefix_with_counts("hello 👋");
        assert_eq!(result.matched_char_count, 7);
        assert_eq!(result.input_char_count, 7); // h-e-l-l-o-space-emoji
    }

    // ==================== Eviction Edge Cases ====================

    #[test]
    fn test_eviction_empty_tree() {
        let tree = Tree::new();

        // Should not panic on empty tree
        tree.evict_tenant_by_size(100);

        let sizes = tree.get_used_size_per_tenant();
        assert!(sizes.is_empty());
    }

    #[test]
    fn test_eviction_zero_max_size() {
        let tree = Tree::new();

        tree.insert_text("hello", "tenant1");
        tree.insert_text("world", "tenant1");

        // Evict with max_size = 0 should remove everything
        tree.evict_tenant_by_size(0);

        let sizes = tree.get_used_size_per_tenant();
        assert!(
            sizes.is_empty() || sizes.values().all(|&v| v == 0),
            "All tenants should be evicted or have zero size"
        );
    }

    #[test]
    fn test_eviction_single_tenant_all_entries() {
        let tree = Tree::new();

        // Insert many entries for single tenant
        for i in 0..100 {
            let text = format!("entry{:03}", i);
            tree.insert_text(&text, "tenant1");
        }

        let initial_size = *tree.get_used_size_per_tenant().get("tenant1").unwrap();
        assert!(initial_size > 50, "Should have significant size");

        // Evict to small size
        tree.evict_tenant_by_size(50);

        let final_size = *tree.get_used_size_per_tenant().get("tenant1").unwrap_or(&0);
        assert!(
            final_size <= 50,
            "Size {} should be <= 50 after eviction",
            final_size
        );
    }

    // ==================== Last Tenant Cache Tests ====================

    #[test]
    fn test_last_tenant_cache_update() {
        let tree = Tree::new();

        // Insert for tenant1
        tree.insert_text("hello", "tenant1");

        // First match should return tenant1
        let (_, tenant) = tree.prefix_match_legacy("hello");
        assert_eq!(tenant, "tenant1");

        // Insert for tenant2 on same path
        tree.insert_text("hello", "tenant2");

        // Match again - should still work (cache or iteration)
        let (matched, _) = tree.prefix_match_legacy("hello");
        assert_eq!(matched, "hello");
    }

    // NOTE: test_stale_cache_after_tenant_removal removed because remove_tenant was removed.
    // TODO: Re-add this test when efficient remove_tenant is implemented.

    // ==================== Consistency Verification Tests ====================

    #[test]
    fn test_char_count_consistency_after_operations() {
        let tree = Tree::new();

        // Helper to verify consistency
        let verify_consistency = |tree: &Tree| {
            let maintained = get_maintained_counts(tree);
            let computed = tree.get_used_size_per_tenant();
            assert_eq!(
                maintained, computed,
                "Maintained counts should match computed counts"
            );
        };

        // Insert phase
        for i in 0..50 {
            tree.insert_text(&format!("prefix{}", i), "tenant1");
            tree.insert_text(&format!("other{}", i), "tenant2");
        }
        verify_consistency(&tree);

        // Overlapping inserts
        for i in 0..25 {
            tree.insert_text(&format!("prefix{}", i), "tenant2");
        }
        verify_consistency(&tree);

        // Eviction
        tree.evict_tenant_by_size(100);
        verify_consistency(&tree);

        // NOTE: Tenant removal test removed because remove_tenant was removed.
        // TODO: Re-add when efficient remove_tenant is implemented.
    }

    #[test]
    fn test_tree_structure_integrity_after_stress() {
        let tree = Arc::new(Tree::new());
        let num_threads = 8;
        let mut handles = vec![];

        for thread_id in 0..num_threads {
            let tree = Arc::clone(&tree);
            let handle = thread::spawn(move || {
                let mut rng = rand::rng();
                let tenant = format!("tenant{}", thread_id);

                for _ in 0..200 {
                    let op: u8 = rng.random_range(0..10);
                    let key = format!("key{}", rng.random_range(0..50));

                    match op {
                        0..=6 => {
                            // Insert (70%)
                            tree.insert_text(&key, &tenant);
                        }
                        7..=8 => {
                            // Match (20%)
                            let _ = tree.prefix_match_legacy(&key);
                        }
                        _ => {
                            // Match with counts (10%)
                            let _ = tree.match_prefix_with_counts(&key);
                        }
                    }
                }
            });
            handles.push(handle);
        }

        for handle in handles {
            handle.join().expect("Thread panicked during stress test");
        }

        // Verify tree is still functional
        let sizes = tree.get_used_size_per_tenant();
        for (tenant, size) in sizes.iter() {
            assert!(*size > 0, "Tenant {} should have positive size", tenant);
        }

        // Verify char count consistency
        let maintained = get_maintained_counts(&tree);
        let computed = tree.get_used_size_per_tenant();
        assert_eq!(
            maintained, computed,
            "Counts should be consistent after stress test"
        );
    }

    // ==================== Boundary Condition Tests ====================

    #[test]
    fn test_very_long_strings() {
        let tree = Tree::new();

        // Create a very long string (10KB)
        let long_string: String = (0..10000)
            .map(|i| ((i % 26) as u8 + b'a') as char)
            .collect();

        tree.insert_text(&long_string, "tenant1");

        let (matched, tenant) = tree.prefix_match_legacy(&long_string);
        assert_eq!(matched.len(), long_string.len());
        assert_eq!(tenant, "tenant1");

        // Partial match of long string
        let partial = &long_string[..5000];
        let (matched, _) = tree.prefix_match_legacy(partial);
        assert_eq!(matched, partial);
    }

    #[test]
    fn test_many_tenants_same_path() {
        let tree = Tree::new();

        // 100 tenants all insert same string
        for i in 0..100 {
            tree.insert_text("shared_path", &format!("tenant{}", i));
        }

        // Match should return one of them
        let (matched, _) = tree.prefix_match_legacy("shared_path");
        assert_eq!(matched, "shared_path");

        // Verify all tenants are tracked
        let sizes = tree.get_used_size_per_tenant();
        assert_eq!(sizes.len(), 100, "Should have 100 tenants");
    }

    #[test]
    fn test_special_characters() {
        let tree = Tree::new();

        // Various special characters
        let test_cases = vec![
            ("hello\nworld", "tenant1"),      // newline
            ("hello\tworld", "tenant2"),      // tab
            ("hello\0world", "tenant3"),      // null byte
            ("hello\u{A0}world", "tenant4"),  // non-breaking space
            ("path/to/file", "tenant5"),      // slashes
            ("query?param=value", "tenant6"), // URL-like
        ];

        for (text, tenant) in &test_cases {
            tree.insert_text(text, tenant);
        }

        for (text, tenant) in &test_cases {
            let (matched, matched_tenant) = tree.prefix_match_legacy(text);
            assert_eq!(matched, *text, "Failed for: {:?}", text);
            assert_eq!(matched_tenant, *tenant);
        }
    }
}