rocksdb-native 0.0.0

package/vendor/rocksdb/tools/block_cache_analyzer/block_cache_pysim.py

@@ -0,0 +1,2000 @@
+#!/usr/bin/env python3
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+
+import gc
+import heapq
+import random
+import sys
+import time
+from collections import OrderedDict
+from os import path
+
+import numpy as np
+
+
+kSampleSize = 64  # The sample size used when performing eviction.
+kMicrosInSecond = 1000000
+kSecondsInMinute = 60
+kSecondsInHour = 3600
+
+
+class TraceRecord:
+    """
+    A trace record represents a block access.
+    It holds the same struct as BlockCacheTraceRecord in
+    trace_replay/block_cache_tracer.h
+    """
+
+    def __init__(
+        self,
+        access_time,
+        block_id,
+        block_type,
+        block_size,
+        cf_id,
+        cf_name,
+        level,
+        fd,
+        caller,
+        no_insert,
+        get_id,
+        key_id,
+        kv_size,
+        is_hit,
+        referenced_key_exist_in_block,
+        num_keys_in_block,
+        table_id,
+        seq_number,
+        block_key_size,
+        key_size,
+        block_offset_in_file,
+        next_access_seq_no,
+    ):
+        self.access_time = access_time
+        self.block_id = block_id
+        self.block_type = block_type
+        self.block_size = block_size + block_key_size
+        self.cf_id = cf_id
+        self.cf_name = cf_name
+        self.level = level
+        self.fd = fd
+        self.caller = caller
+        if no_insert == 1:
+            self.no_insert = True
+        else:
+            self.no_insert = False
+        self.get_id = get_id
+        self.key_id = key_id
+        self.kv_size = kv_size
+        if is_hit == 1:
+            self.is_hit = True
+        else:
+            self.is_hit = False
+        if referenced_key_exist_in_block == 1:
+            self.referenced_key_exist_in_block = True
+        else:
+            self.referenced_key_exist_in_block = False
+        self.num_keys_in_block = num_keys_in_block
+        self.table_id = table_id
+        self.seq_number = seq_number
+        self.block_key_size = block_key_size
+        self.key_size = key_size
+        self.block_offset_in_file = block_offset_in_file
+        self.next_access_seq_no = next_access_seq_no
+
+
+class CacheEntry:
+    """A cache entry stored in the cache."""
+
+    def __init__(
+        self,
+        value_size,
+        cf_id,
+        level,
+        block_type,
+        table_id,
+        access_number,
+        time_s,
+        num_hits=0,
+    ):
+        self.value_size = value_size
+        self.last_access_number = access_number
+        self.num_hits = num_hits
+        self.cf_id = 0
+        self.level = level
+        self.block_type = block_type
+        self.last_access_time = time_s
+        self.insertion_time = time_s
+        self.table_id = table_id
+
+    def __repr__(self):
+        """Debug string."""
+        return "(s={},last={},hits={},cf={},l={},bt={})\n".format(
+            self.value_size,
+            self.last_access_number,
+            self.num_hits,
+            self.cf_id,
+            self.level,
+            self.block_type,
+        )
+
+    def cost_class(self, cost_class_label):
+        if cost_class_label == "table_bt":
+            return "{}-{}".format(self.table_id, self.block_type)
+        elif cost_class_label == "table":
+            return "{}".format(self.table_id)
+        elif cost_class_label == "bt":
+            return "{}".format(self.block_type)
+        elif cost_class_label == "cf":
+            return "{}".format(self.cf_id)
+        elif cost_class_label == "cf_bt":
+            return "{}-{}".format(self.cf_id, self.block_type)
+        elif cost_class_label == "table_level_bt":
+            return "{}-{}-{}".format(self.table_id, self.level, self.block_type)
+        assert False, "Unknown cost class label {}".format(cost_class_label)
+        return None
+
+
+class HashEntry:
+    """A hash entry stored in a hash table."""
+
+    def __init__(self, key, hash, value):
+        self.key = key
+        self.hash = hash
+        self.value = value
+
+    def __repr__(self):
+        return "k={},h={},v=[{}]".format(self.key, self.hash, self.value)
+
+
+class HashTable:
+    """
+    A custom implementation of hash table to support fast random sampling.
+    It is closed hashing and uses chaining to resolve hash conflicts.
+    It grows/shrinks the hash table upon insertion/deletion to support
+    fast lookups and random samplings.
+    """
+
+    def __init__(self):
+        self.initial_size = 32
+        self.table = [None] * self.initial_size
+        self.elements = 0
+
+    def random_sample(self, sample_size):
+        """Randomly sample 'sample_size' hash entries from the table."""
+        samples = []
+        index = random.randint(0, len(self.table) - 1)
+        pos = index
+        # Starting from index, adding hash entries to the sample list until
+        # sample_size is met or we ran out of entries.
+        while True:
+            if self.table[pos] is not None:
+                for i in range(len(self.table[pos])):
+                    if self.table[pos][i] is None:
+                        continue
+                    samples.append(self.table[pos][i])
+                    if len(samples) == sample_size:
+                        break
+            pos += 1
+            pos = pos % len(self.table)
+            if pos == index or len(samples) == sample_size:
+                break
+        assert len(samples) <= sample_size
+        return samples
+
+    def __repr__(self):
+        all_entries = []
+        for i in range(len(self.table)):
+            if self.table[i] is None:
+                continue
+            for j in range(len(self.table[i])):
+                if self.table[i][j] is not None:
+                    all_entries.append(self.table[i][j])
+        return "{}".format(all_entries)
+
+    def values(self):
+        all_values = []
+        for i in range(len(self.table)):
+            if self.table[i] is None:
+                continue
+            for j in range(len(self.table[i])):
+                if self.table[i][j] is not None:
+                    all_values.append(self.table[i][j].value)
+        return all_values
+
+    def __len__(self):
+        return self.elements
+
+    def insert(self, key, hash, value):
+        """
+        Insert a hash entry in the table. Replace the old entry if it already
+        exists.
+        """
+        self.grow()
+        inserted = False
+        index = hash % len(self.table)
+        if self.table[index] is None:
+            self.table[index] = []
+        # Search for the entry first.
+        for i in range(len(self.table[index])):
+            if self.table[index][i] is None:
+                continue
+            if self.table[index][i].hash == hash and self.table[index][i].key == key:
+                # The entry already exists in the table.
+                self.table[index][i] = HashEntry(key, hash, value)
+                return
+
+        # Find an empty slot.
+        for i in range(len(self.table[index])):
+            if self.table[index][i] is None:
+                self.table[index][i] = HashEntry(key, hash, value)
+                inserted = True
+                break
+        if not inserted:
+            self.table[index].append(HashEntry(key, hash, value))
+        self.elements += 1
+
+    def resize(self, new_size):
+        if new_size == len(self.table):
+            return
+        if new_size < self.initial_size:
+            return
+        if self.elements < 100:
+            return
+        new_table = [None] * new_size
+        # Copy 'self.table' to new_table.
+        for i in range(len(self.table)):
+            entries = self.table[i]
+            if entries is None:
+                continue
+            for j in range(len(entries)):
+                if entries[j] is None:
+                    continue
+                index = entries[j].hash % new_size
+                if new_table[index] is None:
+                    new_table[index] = []
+                new_table[index].append(entries[j])
+        self.table = new_table
+        del new_table
+        # Manually call python gc here to free the memory as 'self.table'
+        # might be very large.
+        gc.collect()
+
+    def grow(self):
+        if self.elements < 4 * len(self.table):
+            return
+        new_size = int(len(self.table) * 1.5)
+        self.resize(new_size)
+
+    def delete(self, key, hash):
+        index = hash % len(self.table)
+        deleted = False
+        deleted_entry = None
+        if self.table[index] is None:
+            return
+        for i in range(len(self.table[index])):
+            if (
+                self.table[index][i] is not None
+                and self.table[index][i].hash == hash
+                and self.table[index][i].key == key
+            ):
+                deleted_entry = self.table[index][i]
+                self.table[index][i] = None
+                self.elements -= 1
+                deleted = True
+                break
+        if deleted:
+            self.shrink()
+        return deleted_entry
+
+    def shrink(self):
+        if self.elements * 2 >= len(self.table):
+            return
+        new_size = int(len(self.table) * 0.7)
+        self.resize(new_size)
+
+    def lookup(self, key, hash):
+        index = hash % len(self.table)
+        if self.table[index] is None:
+            return None
+        for i in range(len(self.table[index])):
+            if (
+                self.table[index][i] is not None
+                and self.table[index][i].hash == hash
+                and self.table[index][i].key == key
+            ):
+                return self.table[index][i].value
+        return None
+
+
+class MissRatioStats:
+    def __init__(self, time_unit):
+        self.num_misses = 0
+        self.num_accesses = 0
+        self.time_unit = time_unit
+        self.time_misses = {}
+        self.time_miss_bytes = {}
+        self.time_accesses = {}
+
+    def update_metrics(self, access_time, is_hit, miss_bytes):
+        access_time /= kMicrosInSecond * self.time_unit
+        self.num_accesses += 1
+        if access_time not in self.time_accesses:
+            self.time_accesses[access_time] = 0
+        self.time_accesses[access_time] += 1
+        if not is_hit:
+            self.num_misses += 1
+            if access_time not in self.time_misses:
+                self.time_misses[access_time] = 0
+                self.time_miss_bytes[access_time] = 0
+            self.time_misses[access_time] += 1
+            self.time_miss_bytes[access_time] += miss_bytes
+
+    def reset_counter(self):
+        self.num_misses = 0
+        self.num_accesses = 0
+        self.time_miss_bytes.clear()
+        self.time_misses.clear()
+        self.time_accesses.clear()
+
+    def compute_miss_bytes(self):
+        miss_bytes = []
+        for at in self.time_miss_bytes:
+            miss_bytes.append(self.time_miss_bytes[at])
+        miss_bytes = sorted(miss_bytes)
+        avg_miss_bytes = 0
+        p95_miss_bytes = 0
+        for i in range(len(miss_bytes)):
+            avg_miss_bytes += float(miss_bytes[i]) / float(len(miss_bytes))
+
+        p95_index = min(int(0.95 * float(len(miss_bytes))), len(miss_bytes) - 1)
+        p95_miss_bytes = miss_bytes[p95_index]
+        return avg_miss_bytes, p95_miss_bytes
+
+    def miss_ratio(self):
+        return float(self.num_misses) * 100.0 / float(self.num_accesses)
+
+    def write_miss_timeline(
+        self, cache_type, cache_size, target_cf_name, result_dir, start, end
+    ):
+        start /= kMicrosInSecond * self.time_unit
+        end /= kMicrosInSecond * self.time_unit
+        header_file_path = "{}/header-ml-miss-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        if not path.exists(header_file_path):
+            with open(header_file_path, "w+") as header_file:
+                header = "time"
+                for trace_time in range(start, end):
+                    header += ",{}".format(trace_time)
+                header_file.write(header + "\n")
+        file_path = "{}/data-ml-miss-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        with open(file_path, "w+") as file:
+            row = "{}".format(cache_type)
+            for trace_time in range(start, end):
+                row += ",{}".format(self.time_misses.get(trace_time, 0))
+            file.write(row + "\n")
+
+    def write_miss_ratio_timeline(
+        self, cache_type, cache_size, target_cf_name, result_dir, start, end
+    ):
+        start /= kMicrosInSecond * self.time_unit
+        end /= kMicrosInSecond * self.time_unit
+        header_file_path = "{}/header-ml-miss-ratio-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        if not path.exists(header_file_path):
+            with open(header_file_path, "w+") as header_file:
+                header = "time"
+                for trace_time in range(start, end):
+                    header += ",{}".format(trace_time)
+                header_file.write(header + "\n")
+        file_path = "{}/data-ml-miss-ratio-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        with open(file_path, "w+") as file:
+            row = "{}".format(cache_type)
+            for trace_time in range(start, end):
+                naccesses = self.time_accesses.get(trace_time, 0)
+                miss_ratio = 0
+                if naccesses > 0:
+                    miss_ratio = float(
+                        self.time_misses.get(trace_time, 0) * 100.0
+                    ) / float(naccesses)
+                row += ",{0:.2f}".format(miss_ratio)
+            file.write(row + "\n")
+
+
+class PolicyStats:
+    def __init__(self, time_unit, policies):
+        self.time_selected_polices = {}
+        self.time_accesses = {}
+        self.policy_names = {}
+        self.time_unit = time_unit
+        for i in range(len(policies)):
+            self.policy_names[i] = policies[i].policy_name()
+
+    def update_metrics(self, access_time, selected_policy):
+        access_time /= kMicrosInSecond * self.time_unit
+        if access_time not in self.time_accesses:
+            self.time_accesses[access_time] = 0
+        self.time_accesses[access_time] += 1
+        if access_time not in self.time_selected_polices:
+            self.time_selected_polices[access_time] = {}
+        policy_name = self.policy_names[selected_policy]
+        if policy_name not in self.time_selected_polices[access_time]:
+            self.time_selected_polices[access_time][policy_name] = 0
+        self.time_selected_polices[access_time][policy_name] += 1
+
+    def write_policy_timeline(
+        self, cache_type, cache_size, target_cf_name, result_dir, start, end
+    ):
+        start /= kMicrosInSecond * self.time_unit
+        end /= kMicrosInSecond * self.time_unit
+        header_file_path = "{}/header-ml-policy-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        if not path.exists(header_file_path):
+            with open(header_file_path, "w+") as header_file:
+                header = "time"
+                for trace_time in range(start, end):
+                    header += ",{}".format(trace_time)
+                header_file.write(header + "\n")
+        file_path = "{}/data-ml-policy-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        with open(file_path, "w+") as file:
+            for policy in self.policy_names:
+                policy_name = self.policy_names[policy]
+                row = "{}-{}".format(cache_type, policy_name)
+                for trace_time in range(start, end):
+                    row += ",{}".format(
+                        self.time_selected_polices.get(trace_time, {}).get(
+                            policy_name, 0
+                        )
+                    )
+                file.write(row + "\n")
+
+    def write_policy_ratio_timeline(
+        self, cache_type, cache_size, target_cf_name, file_path, start, end
+    ):
+        start /= kMicrosInSecond * self.time_unit
+        end /= kMicrosInSecond * self.time_unit
+        header_file_path = "{}/header-ml-policy-ratio-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        if not path.exists(header_file_path):
+            with open(header_file_path, "w+") as header_file:
+                header = "time"
+                for trace_time in range(start, end):
+                    header += ",{}".format(trace_time)
+                header_file.write(header + "\n")
+        file_path = "{}/data-ml-policy-ratio-timeline-{}-{}-{}-{}".format(
+            result_dir, self.time_unit, cache_type, cache_size, target_cf_name
+        )
+        with open(file_path, "w+") as file:
+            for policy in self.policy_names:
+                policy_name = self.policy_names[policy]
+                row = "{}-{}".format(cache_type, policy_name)
+                for trace_time in range(start, end):
+                    naccesses = self.time_accesses.get(trace_time, 0)
+                    ratio = 0
+                    if naccesses > 0:
+                        ratio = float(
+                            self.time_selected_polices.get(trace_time, {}).get(
+                                policy_name, 0
+                            )
+                            * 100.0
+                        ) / float(naccesses)
+                    row += ",{0:.2f}".format(ratio)
+                file.write(row + "\n")
+
+
+class Policy(object):
+    """
+    A policy maintains a set of evicted keys. It returns a reward of one to
+    itself if it has not evicted a missing key. Otherwise, it gives itself 0
+    reward.
+    """
+
+    def __init__(self):
+        self.evicted_keys = {}
+
+    def evict(self, key, max_size):
+        self.evicted_keys[key] = 0
+
+    def delete(self, key):
+        self.evicted_keys.pop(key, None)
+
+    def prioritize_samples(self, samples, auxilliary_info):
+        raise NotImplementedError
+
+    def policy_name(self):
+        raise NotImplementedError
+
+    def generate_reward(self, key):
+        if key in self.evicted_keys:
+            return 0
+        return 1
+
+
+class LRUPolicy(Policy):
+    def prioritize_samples(self, samples, auxilliary_info):
+        return sorted(
+            samples,
+            cmp=lambda e1, e2: e1.value.last_access_number
+            - e2.value.last_access_number,
+        )
+
+    def policy_name(self):
+        return "lru"
+
+
+class MRUPolicy(Policy):
+    def prioritize_samples(self, samples, auxilliary_info):
+        return sorted(
+            samples,
+            cmp=lambda e1, e2: e2.value.last_access_number
+            - e1.value.last_access_number,
+        )
+
+    def policy_name(self):
+        return "mru"
+
+
+class LFUPolicy(Policy):
+    def prioritize_samples(self, samples, auxilliary_info):
+        return sorted(samples, cmp=lambda e1, e2: e1.value.num_hits - e2.value.num_hits)
+
+    def policy_name(self):
+        return "lfu"
+
+
+class HyperbolicPolicy(Policy):
+    """
+    An implementation of Hyperbolic caching.
+
+    Aaron Blankstein, Siddhartha Sen, and Michael J. Freedman. 2017.
+    Hyperbolic caching: flexible caching for web applications. In Proceedings
+    of the 2017 USENIX Conference on Usenix Annual Technical Conference
+    (USENIX ATC '17). USENIX Association, Berkeley, CA, USA, 499-511.
+    """
+
+    def compare(self, e1, e2, now):
+        e1_duration = max(0, (now - e1.value.insertion_time) / kMicrosInSecond) * float(
+            e1.value.value_size
+        )
+        e2_duration = max(0, (now - e2.value.insertion_time) / kMicrosInSecond) * float(
+            e2.value.value_size
+        )
+        if e1_duration == e2_duration:
+            return e1.value.num_hits - e2.value.num_hits
+        if e1_duration == 0:
+            return 1
+        if e2_duration == 0:
+            return 1
+        diff = (float(e1.value.num_hits) / (float(e1_duration))) - (
+            float(e2.value.num_hits) / float(e2_duration)
+        )
+        if diff == 0:
+            return 0
+        elif diff > 0:
+            return 1
+        else:
+            return -1
+
+    def prioritize_samples(self, samples, auxilliary_info):
+        assert len(auxilliary_info) == 3
+        now = auxilliary_info[0]
+        return sorted(samples, cmp=lambda e1, e2: self.compare(e1, e2, now))
+
+    def policy_name(self):
+        return "hb"
+
+
+class CostClassPolicy(Policy):
+    """
+    We calculate the hit density of a cost class as
+    number of hits / total size in cache * average duration in the cache.
+
+    An entry has a higher priority if its class's hit density is higher.
+    """
+
+    def compare(self, e1, e2, now, cost_classes, cost_class_label):
+        e1_class = e1.value.cost_class(cost_class_label)
+        e2_class = e2.value.cost_class(cost_class_label)
+
+        assert e1_class in cost_classes
+        assert e2_class in cost_classes
+
+        e1_entry = cost_classes[e1_class]
+        e2_entry = cost_classes[e2_class]
+        e1_density = e1_entry.density(now)
+        e2_density = e2_entry.density(now)
+        e1_hits = cost_classes[e1_class].hits
+        e2_hits = cost_classes[e2_class].hits
+
+        if e1_density == e2_density:
+            return e1_hits - e2_hits
+
+        if e1_entry.num_entries_in_cache == 0:
+            return -1
+        if e2_entry.num_entries_in_cache == 0:
+            return 1
+
+        if e1_density == 0:
+            return 1
+        if e2_density == 0:
+            return -1
+        diff = (float(e1_hits) / float(e1_density)) - (
+            float(e2_hits) / float(e2_density)
+        )
+        if diff == 0:
+            return 0
+        elif diff > 0:
+            return 1
+        else:
+            return -1
+
+    def prioritize_samples(self, samples, auxilliary_info):
+        assert len(auxilliary_info) == 3
+        now = auxilliary_info[0]
+        cost_classes = auxilliary_info[1]
+        cost_class_label = auxilliary_info[2]
+        return sorted(
+            samples,
+            cmp=lambda e1, e2: self.compare(
+                e1, e2, now, cost_classes, cost_class_label
+            ),
+        )
+
+    def policy_name(self):
+        return "cc"
+
+
+class Cache(object):
+    """
+    This is the base class for the implementations of alternative cache
+    replacement policies.
+    """
+
+    def __init__(self, cache_size, enable_cache_row_key):
+        self.cache_size = cache_size
+        self.used_size = 0
+        self.per_second_miss_ratio_stats = MissRatioStats(1)
+        self.miss_ratio_stats = MissRatioStats(kSecondsInMinute)
+        self.per_hour_miss_ratio_stats = MissRatioStats(kSecondsInHour)
+        # 0: disabled. 1: enabled. Insert both row and the refereneced data block.
+        # 2: enabled. Insert only the row but NOT the referenced data block.
+        self.enable_cache_row_key = enable_cache_row_key
+        self.get_id_row_key_map = {}
+        self.max_seen_get_id = 0
+        self.retain_get_id_range = 100000
+
+    def block_key(self, trace_record):
+        return "b{}".format(trace_record.block_id)
+
+    def row_key(self, trace_record):
+        return "g{}-{}".format(trace_record.fd, trace_record.key_id)
+
+    def _lookup(self, trace_record, key, hash):
+        """
+        Look up the key in the cache.
+        Returns true upon a cache hit, false otherwise.
+        """
+        raise NotImplementedError
+
+    def _evict(self, trace_record, key, hash, value_size):
+        """
+        Evict entries in the cache until there is enough room to insert the new
+        entry with 'value_size'.
+        """
+        raise NotImplementedError
+
+    def _insert(self, trace_record, key, hash, value_size):
+        """
+        Insert the new entry into the cache.
+        """
+        raise NotImplementedError
+
+    def _should_admit(self, trace_record, key, hash, value_size):
+        """
+        A custom admission policy to decide whether we should admit the new
+        entry upon a cache miss.
+        Returns true if the new entry should be admitted, false otherwise.
+        """
+        raise NotImplementedError
+
+    def cache_name(self):
+        """
+        The name of the replacement policy.
+        """
+        raise NotImplementedError
+
+    def is_ml_cache(self):
+        return False
+
+    def _update_stats(self, access_time, is_hit, miss_bytes):
+        self.per_second_miss_ratio_stats.update_metrics(access_time, is_hit, miss_bytes)
+        self.miss_ratio_stats.update_metrics(access_time, is_hit, miss_bytes)
+        self.per_hour_miss_ratio_stats.update_metrics(access_time, is_hit, miss_bytes)
+
+    def access(self, trace_record):
+        """
+        Access a trace record. The simulator calls this function to access a
+        trace record.
+        """
+        assert self.used_size <= self.cache_size
+        if (
+            self.enable_cache_row_key > 0
+            and trace_record.caller == 1
+            and trace_record.key_id != 0
+            and trace_record.get_id != 0
+        ):
+            # This is a get request.
+            self._access_row(trace_record)
+            return
+        is_hit = self._access_kv(
+            trace_record,
+            self.block_key(trace_record),
+            trace_record.block_id,
+            trace_record.block_size,
+            trace_record.no_insert,
+        )
+        self._update_stats(
+            trace_record.access_time, is_hit=is_hit, miss_bytes=trace_record.block_size
+        )
+
+    def _access_row(self, trace_record):
+        row_key = self.row_key(trace_record)
+        self.max_seen_get_id = max(self.max_seen_get_id, trace_record.get_id)
+        self.get_id_row_key_map.pop(
+            self.max_seen_get_id - self.retain_get_id_range, None
+        )
+        if trace_record.get_id not in self.get_id_row_key_map:
+            self.get_id_row_key_map[trace_record.get_id] = {}
+            self.get_id_row_key_map[trace_record.get_id]["h"] = False
+        if self.get_id_row_key_map[trace_record.get_id]["h"]:
+            # We treat future accesses as hits since this get request
+            # completes.
+            # print("row hit 1")
+            self._update_stats(trace_record.access_time, is_hit=True, miss_bytes=0)
+            return
+        if row_key not in self.get_id_row_key_map[trace_record.get_id]:
+            # First time seen this key.
+            is_hit = self._access_kv(
+                trace_record,
+                key=row_key,
+                hash=trace_record.key_id,
+                value_size=trace_record.kv_size,
+                no_insert=False,
+            )
+            inserted = False
+            if trace_record.kv_size > 0:
+                inserted = True
+            self.get_id_row_key_map[trace_record.get_id][row_key] = inserted
+            self.get_id_row_key_map[trace_record.get_id]["h"] = is_hit
+        if self.get_id_row_key_map[trace_record.get_id]["h"]:
+            # We treat future accesses as hits since this get request
+            # completes.
+            # print("row hit 2")
+            self._update_stats(trace_record.access_time, is_hit=True, miss_bytes=0)
+            return
+        # Access its blocks.
+        no_insert = trace_record.no_insert
+        if (
+            self.enable_cache_row_key == 2
+            and trace_record.kv_size > 0
+            and trace_record.block_type == 9
+        ):
+            no_insert = True
+        is_hit = self._access_kv(
+            trace_record,
+            key=self.block_key(trace_record),
+            hash=trace_record.block_id,
+            value_size=trace_record.block_size,
+            no_insert=no_insert,
+        )
+        self._update_stats(
+            trace_record.access_time, is_hit, miss_bytes=trace_record.block_size
+        )
+        if (
+            trace_record.kv_size > 0
+            and not self.get_id_row_key_map[trace_record.get_id][row_key]
+        ):
+            # Insert the row key-value pair.
+            self._access_kv(
+                trace_record,
+                key=row_key,
+                hash=trace_record.key_id,
+                value_size=trace_record.kv_size,
+                no_insert=False,
+            )
+            # Mark as inserted.
+            self.get_id_row_key_map[trace_record.get_id][row_key] = True
+
+    def _access_kv(self, trace_record, key, hash, value_size, no_insert):
+        # Sanity checks.
+        assert self.used_size <= self.cache_size
+        if self._lookup(trace_record, key, hash):
+            # A cache hit.
+            return True
+        if no_insert or value_size <= 0:
+            return False
+        # A cache miss.
+        if value_size > self.cache_size:
+            # The block is too large to fit into the cache.
+            return False
+        self._evict(trace_record, key, hash, value_size)
+        if self._should_admit(trace_record, key, hash, value_size):
+            self._insert(trace_record, key, hash, value_size)
+            self.used_size += value_size
+        return False
+
+
+class CostClassEntry:
+    """
+    A cost class maintains aggregated statistics of cached entries in a class.
+    For example, we may define block type as a class. Then, cached blocks of the
+    same type will share one cost class entry.
+    """
+
+    def __init__(self):
+        self.hits = 0
+        self.num_entries_in_cache = 0
+        self.size_in_cache = 0
+        self.sum_insertion_times = 0
+        self.sum_last_access_time = 0
+
+    def insert(self, trace_record, key, value_size):
+        self.size_in_cache += value_size
+        self.num_entries_in_cache += 1
+        self.sum_insertion_times += trace_record.access_time / kMicrosInSecond
+        self.sum_last_access_time += trace_record.access_time / kMicrosInSecond
+
+    def remove(self, insertion_time, last_access_time, key, value_size, num_hits):
+        self.hits -= num_hits
+        self.num_entries_in_cache -= 1
+        self.sum_insertion_times -= insertion_time / kMicrosInSecond
+        self.size_in_cache -= value_size
+        self.sum_last_access_time -= last_access_time / kMicrosInSecond
+
+    def update_on_hit(self, trace_record, last_access_time):
+        self.hits += 1
+        self.sum_last_access_time -= last_access_time / kMicrosInSecond
+        self.sum_last_access_time += trace_record.access_time / kMicrosInSecond
+
+    def avg_lifetime_in_cache(self, now):
+        avg_insertion_time = self.sum_insertion_times / self.num_entries_in_cache
+        return now / kMicrosInSecond - avg_insertion_time
+
+    def avg_last_access_time(self):
+        if self.num_entries_in_cache == 0:
+            return 0
+        return float(self.sum_last_access_time) / float(self.num_entries_in_cache)
+
+    def avg_size(self):
+        if self.num_entries_in_cache == 0:
+            return 0
+        return float(self.sum_last_access_time) / float(self.num_entries_in_cache)
+
+    def density(self, now):
+        avg_insertion_time = self.sum_insertion_times / self.num_entries_in_cache
+        in_cache_duration = now / kMicrosInSecond - avg_insertion_time
+        return self.size_in_cache * in_cache_duration
+
+
+class MLCache(Cache):
+    """
+    MLCache is the base class for implementations of alternative replacement
+    policies using reinforcement learning.
+    """
+
+    def __init__(self, cache_size, enable_cache_row_key, policies, cost_class_label):
+        super(MLCache, self).__init__(cache_size, enable_cache_row_key)
+        self.table = HashTable()
+        self.policy_stats = PolicyStats(kSecondsInMinute, policies)
+        self.per_hour_policy_stats = PolicyStats(kSecondsInHour, policies)
+        self.policies = policies
+        self.cost_classes = {}
+        self.cost_class_label = cost_class_label
+
+    def is_ml_cache(self):
+        return True
+
+    def _lookup(self, trace_record, key, hash):
+        value = self.table.lookup(key, hash)
+        if value is not None:
+            # Update the entry's cost class statistics.
+            if self.cost_class_label is not None:
+                cost_class = value.cost_class(self.cost_class_label)
+                assert cost_class in self.cost_classes
+                self.cost_classes[cost_class].update_on_hit(
+                    trace_record, value.last_access_time
+                )
+            # Update the entry's last access time.
+            self.table.insert(
+                key,
+                hash,
+                CacheEntry(
+                    value_size=value.value_size,
+                    cf_id=value.cf_id,
+                    level=value.level,
+                    block_type=value.block_type,
+                    table_id=value.table_id,
+                    access_number=self.miss_ratio_stats.num_accesses,
+                    time_s=trace_record.access_time,
+                    num_hits=value.num_hits + 1,
+                ),
+            )
+            return True
+        return False
+
+    def _evict(self, trace_record, key, hash, value_size):
+        # Select a policy, random sample kSampleSize keys from the cache, then
+        # evict keys in the sample set until we have enough room for the new
+        # entry.
+        policy_index = self._select_policy(trace_record, key)
+        assert policy_index < len(self.policies) and policy_index >= 0
+        self.policies[policy_index].delete(key)
+        self.policy_stats.update_metrics(trace_record.access_time, policy_index)
+        self.per_hour_policy_stats.update_metrics(
+            trace_record.access_time, policy_index
+        )
+        while self.used_size + value_size > self.cache_size:
+            # Randomly sample n entries.
+            samples = self.table.random_sample(kSampleSize)
+            samples = self.policies[policy_index].prioritize_samples(
+                samples,
+                [trace_record.access_time, self.cost_classes, self.cost_class_label],
+            )
+            for hash_entry in samples:
+                assert self.table.delete(hash_entry.key, hash_entry.hash) is not None
+                self.used_size -= hash_entry.value.value_size
+                self.policies[policy_index].evict(
+                    key=hash_entry.key, max_size=self.table.elements
+                )
+                # Update the entry's cost class statistics.
+                if self.cost_class_label is not None:
+                    cost_class = hash_entry.value.cost_class(self.cost_class_label)
+                    assert cost_class in self.cost_classes
+                    self.cost_classes[cost_class].remove(
+                        hash_entry.value.insertion_time,
+                        hash_entry.value.last_access_time,
+                        key,
+                        hash_entry.value.value_size,
+                        hash_entry.value.num_hits,
+                    )
+                if self.used_size + value_size <= self.cache_size:
+                    break
+
+    def _insert(self, trace_record, key, hash, value_size):
+        assert self.used_size + value_size <= self.cache_size
+        entry = CacheEntry(
+            value_size,
+            trace_record.cf_id,
+            trace_record.level,
+            trace_record.block_type,
+            trace_record.table_id,
+            self.miss_ratio_stats.num_accesses,
+            trace_record.access_time,
+        )
+        # Update the entry's cost class statistics.
+        if self.cost_class_label is not None:
+            cost_class = entry.cost_class(self.cost_class_label)
+            if cost_class not in self.cost_classes:
+                self.cost_classes[cost_class] = CostClassEntry()
+            self.cost_classes[cost_class].insert(trace_record, key, value_size)
+        self.table.insert(key, hash, entry)
+
+    def _should_admit(self, trace_record, key, hash, value_size):
+        return True
+
+    def _select_policy(self, trace_record, key):
+        raise NotImplementedError
+
+
+class ThompsonSamplingCache(MLCache):
+    """
+    An implementation of Thompson Sampling for the Bernoulli Bandit.
+
+    Daniel J. Russo, Benjamin Van Roy, Abbas Kazerouni, Ian Osband,
+    and Zheng Wen. 2018. A Tutorial on Thompson Sampling. Found.
+    Trends Mach. Learn. 11, 1 (July 2018), 1-96.
+    DOI: https://doi.org/10.1561/2200000070
+    """
+
+    def __init__(
+        self,
+        cache_size,
+        enable_cache_row_key,
+        policies,
+        cost_class_label,
+        init_a=1,
+        init_b=1,
+    ):
+        super(ThompsonSamplingCache, self).__init__(
+            cache_size, enable_cache_row_key, policies, cost_class_label
+        )
+        self._as = {}
+        self._bs = {}
+        for _i in range(len(policies)):
+            self._as = [init_a] * len(self.policies)
+            self._bs = [init_b] * len(self.policies)
+
+    def _select_policy(self, trace_record, key):
+        if len(self.policies) == 1:
+            return 0
+        samples = [
+            np.random.beta(self._as[x], self._bs[x]) for x in range(len(self.policies))
+        ]
+        selected_policy = max(range(len(self.policies)), key=lambda x: samples[x])
+        reward = self.policies[selected_policy].generate_reward(key)
+        assert reward <= 1 and reward >= 0
+        self._as[selected_policy] += reward
+        self._bs[selected_policy] += 1 - reward
+        return selected_policy
+
+    def cache_name(self):
+        if self.enable_cache_row_key:
+            return "Hybrid ThompsonSampling with cost class {} (ts_hybrid)".format(
+                self.cost_class_label
+            )
+        return "ThompsonSampling with cost class {} (ts)".format(self.cost_class_label)
+
+
+class LinUCBCache(MLCache):
+    """
+    An implementation of LinUCB with disjoint linear models.
+
+    Lihong Li, Wei Chu, John Langford, and Robert E. Schapire. 2010.
+    A contextual-bandit approach to personalized news article recommendation.
+    In Proceedings of the 19th international conference on World wide web
+    (WWW '10). ACM, New York, NY, USA, 661-670.
+    DOI=http://dx.doi.org/10.1145/1772690.1772758
+    """
+
+    def __init__(self, cache_size, enable_cache_row_key, policies, cost_class_label):
+        super(LinUCBCache, self).__init__(
+            cache_size, enable_cache_row_key, policies, cost_class_label
+        )
+        self.nfeatures = 4  # Block type, level, cf.
+        self.th = np.zeros((len(self.policies), self.nfeatures))
+        self.eps = 0.2
+        self.b = np.zeros_like(self.th)
+        self.A = np.zeros((len(self.policies), self.nfeatures, self.nfeatures))
+        self.A_inv = np.zeros((len(self.policies), self.nfeatures, self.nfeatures))
+        for i in range(len(self.policies)):
+            self.A[i] = np.identity(self.nfeatures)
+        self.th_hat = np.zeros_like(self.th)
+        self.p = np.zeros(len(self.policies))
+        self.alph = 0.2
+
+    def _select_policy(self, trace_record, key):
+        if len(self.policies) == 1:
+            return 0
+        x_i = np.zeros(self.nfeatures)  # The current context vector
+        x_i[0] = trace_record.block_type
+        x_i[1] = trace_record.level
+        x_i[2] = trace_record.cf_id
+        p = np.zeros(len(self.policies))
+        for a in range(len(self.policies)):
+            self.th_hat[a] = self.A_inv[a].dot(self.b[a])
+            ta = x_i.dot(self.A_inv[a]).dot(x_i)
+            a_upper_ci = self.alph * np.sqrt(ta)
+            a_mean = self.th_hat[a].dot(x_i)
+            p[a] = a_mean + a_upper_ci
+        p = p + (np.random.random(len(p)) * 0.000001)
+        selected_policy = p.argmax()
+        reward = self.policies[selected_policy].generate_reward(key)
+        assert reward <= 1 and reward >= 0
+        self.A[selected_policy] += np.outer(x_i, x_i)
+        self.b[selected_policy] += reward * x_i
+        self.A_inv[selected_policy] = np.linalg.inv(self.A[selected_policy])
+        del x_i
+        return selected_policy
+
+    def cache_name(self):
+        if self.enable_cache_row_key:
+            return "Hybrid LinUCB with cost class {} (linucb_hybrid)".format(
+                self.cost_class_label
+            )
+        return "LinUCB with cost class {} (linucb)".format(self.cost_class_label)
+
+
+class OPTCacheEntry:
+    """
+    A cache entry for the OPT algorithm. The entries are sorted based on its
+    next access sequence number in reverse order, i.e., the entry which next
+    access is the furthest in the future is ordered before other entries.
+    """
+
+    def __init__(self, key, next_access_seq_no, value_size):
+        self.key = key
+        self.next_access_seq_no = next_access_seq_no
+        self.value_size = value_size
+        self.is_removed = False
+
+    def __cmp__(self, other):
+        if other.next_access_seq_no != self.next_access_seq_no:
+            return other.next_access_seq_no - self.next_access_seq_no
+        return self.value_size - other.value_size
+
+    def __repr__(self):
+        return "({} {} {} {})".format(
+            self.key, self.next_access_seq_no, self.value_size, self.is_removed
+        )
+
+
+class PQTable:
+    """
+    A hash table with a priority queue.
+    """
+
+    def __init__(self):
+        # A list of entries arranged in a heap sorted based on the entry custom
+        # implementation of __cmp__
+        self.pq = []
+        self.table = {}
+
+    def pqinsert(self, entry):
+        "Add a new key or update the priority of an existing key"
+        # Remove the entry from the table first.
+        removed_entry = self.table.pop(entry.key, None)
+        if removed_entry:
+            # Mark as removed since there is no 'remove' API in heappq.
+            # Instead, an entry in pq is removed lazily when calling pop.
+            removed_entry.is_removed = True
+        self.table[entry.key] = entry
+        heapq.heappush(self.pq, entry)
+        return removed_entry
+
+    def pqpop(self):
+        while self.pq:
+            entry = heapq.heappop(self.pq)
+            if not entry.is_removed:
+                del self.table[entry.key]
+                return entry
+        return None
+
+    def pqpeek(self):
+        while self.pq:
+            entry = self.pq[0]
+            if not entry.is_removed:
+                return entry
+            heapq.heappop(self.pq)
+        return
+
+    def __contains__(self, k):
+        return k in self.table
+
+    def __getitem__(self, k):
+        return self.table[k]
+
+    def __len__(self):
+        return len(self.table)
+
+    def values(self):
+        return self.table.values()
+
+
+class OPTCache(Cache):
+    """
+    An implementation of the Belady MIN algorithm. OPTCache evicts an entry
+    in the cache whose next access occurs furthest in the future.
+
+    Note that Belady MIN algorithm is optimal assuming all blocks having the
+    same size and a missing entry will be inserted in the cache.
+    These are NOT true for the block cache trace since blocks have different
+    sizes and we may not insert a block into the cache upon a cache miss.
+    However, it is still useful to serve as a "theoretical upper bound" on the
+    lowest miss ratio we can achieve given a cache size.
+
+    L. A. Belady. 1966. A Study of Replacement Algorithms for a
+    Virtual-storage Computer. IBM Syst. J. 5, 2 (June 1966), 78-101.
+    DOI=http://dx.doi.org/10.1147/sj.52.0078
+    """
+
+    def __init__(self, cache_size):
+        super(OPTCache, self).__init__(cache_size, enable_cache_row_key=0)
+        self.table = PQTable()
+
+    def _lookup(self, trace_record, key, hash):
+        if key not in self.table:
+            return False
+        # A cache hit. Update its next access time.
+        assert (
+            self.table.pqinsert(
+                OPTCacheEntry(
+                    key, trace_record.next_access_seq_no, self.table[key].value_size
+                )
+            )
+            is not None
+        )
+        return True
+
+    def _evict(self, trace_record, key, hash, value_size):
+        while self.used_size + value_size > self.cache_size:
+            evict_entry = self.table.pqpop()
+            assert evict_entry is not None
+            self.used_size -= evict_entry.value_size
+
+    def _insert(self, trace_record, key, hash, value_size):
+        assert (
+            self.table.pqinsert(
+                OPTCacheEntry(key, trace_record.next_access_seq_no, value_size)
+            )
+            is None
+        )
+
+    def _should_admit(self, trace_record, key, hash, value_size):
+        return True
+
+    def cache_name(self):
+        return "Belady MIN (opt)"
+
+
+class GDSizeEntry:
+    """
+    A cache entry for the greedy dual size replacement policy.
+    """
+
+    def __init__(self, key, value_size, priority):
+        self.key = key
+        self.value_size = value_size
+        self.priority = priority
+        self.is_removed = False
+
+    def __cmp__(self, other):
+        if other.priority != self.priority:
+            return self.priority - other.priority
+        return self.value_size - other.value_size
+
+    def __repr__(self):
+        return "({} {} {} {})".format(
+            self.key, self.next_access_seq_no, self.value_size, self.is_removed
+        )
+
+
+class GDSizeCache(Cache):
+    """
+    An implementation of the greedy dual size algorithm.
+    We define cost as an entry's size.
+
+    See https://www.usenix.org/legacy/publications/library/proceedings/usits97/full_papers/cao/cao_html/node8.html
+    and N. Young. The k-server dual and loose competitiveness for paging.
+    Algorithmica,June 1994, vol. 11,(no.6):525-41.
+    Rewritten version of ''On-line caching as cache size varies'',
+    in The 2nd Annual ACM-SIAM Symposium on Discrete Algorithms, 241-250, 1991.
+    """
+
+    def __init__(self, cache_size, enable_cache_row_key):
+        super(GDSizeCache, self).__init__(cache_size, enable_cache_row_key)
+        self.table = PQTable()
+        self.L = 0.0
+
+    def cache_name(self):
+        if self.enable_cache_row_key:
+            return "Hybrid GreedyDualSize (gdsize_hybrid)"
+        return "GreedyDualSize (gdsize)"
+
+    def _lookup(self, trace_record, key, hash):
+        if key not in self.table:
+            return False
+        # A cache hit. Update its priority.
+        entry = self.table[key]
+        assert (
+            self.table.pqinsert(
+                GDSizeEntry(key, entry.value_size, self.L + entry.value_size)
+            )
+            is not None
+        )
+        return True
+
+    def _evict(self, trace_record, key, hash, value_size):
+        while self.used_size + value_size > self.cache_size:
+            evict_entry = self.table.pqpop()
+            assert evict_entry is not None
+            self.L = evict_entry.priority
+            self.used_size -= evict_entry.value_size
+
+    def _insert(self, trace_record, key, hash, value_size):
+        assert (
+            self.table.pqinsert(GDSizeEntry(key, value_size, self.L + value_size))
+            is None
+        )
+
+    def _should_admit(self, trace_record, key, hash, value_size):
+        return True
+
+
+class Deque(object):
+    """A Deque class facilitates the implementation of LRU and ARC."""
+
+    def __init__(self):
+        self.od = OrderedDict()
+
+    def appendleft(self, k):
+        if k in self.od:
+            del self.od[k]
+        self.od[k] = None
+
+    def pop(self):
+        item = self.od.popitem(last=False) if self.od else None
+        if item is not None:
+            return item[0]
+        return None
+
+    def remove(self, k):
+        del self.od[k]
+
+    def __len__(self):
+        return len(self.od)
+
+    def __contains__(self, k):
+        return k in self.od
+
+    def __iter__(self):
+        return reversed(self.od)
+
+    def __repr__(self):
+        return "Deque(%r)" % (list(self),)
+
+
+class ARCCache(Cache):
+    """
+    An implementation of ARC. ARC assumes that all blocks are having the
+    same size. The size of index and filter blocks are variable. To accommodate
+    this, we modified ARC as follows:
+    1) We use 16 KB as the average block size and calculate the number of blocks
+       (c) in the cache.
+    2) When we insert an entry, the cache evicts entries in both t1 and t2
+       queues until it has enough space for the new entry. This also requires
+       modification of the algorithm to maintain a maximum of 2*c blocks.
+
+    Nimrod Megiddo and Dharmendra S. Modha. 2003. ARC: A Self-Tuning, Low
+    Overhead Replacement Cache. In Proceedings of the 2nd USENIX Conference on
+    File and Storage Technologies (FAST '03). USENIX Association, Berkeley, CA,
+    USA, 115-130.
+    """
+
+    def __init__(self, cache_size, enable_cache_row_key):
+        super(ARCCache, self).__init__(cache_size, enable_cache_row_key)
+        self.table = {}
+        self.c = cache_size / 16 * 1024  # Number of elements in the cache.
+        self.p = 0  # Target size for the list T1
+        # L1: only once recently
+        self.t1 = Deque()  # T1: recent cache entries
+        self.b1 = Deque()  # B1: ghost entries recently evicted from the T1 cache
+        # L2: at least twice recently
+        self.t2 = Deque()  # T2: frequent entries
+        self.b2 = Deque()  # B2: ghost entries recently evicted from the T2 cache
+
+    def _replace(self, key, value_size):
+        while self.used_size + value_size > self.cache_size:
+            if self.t1 and ((key in self.b2) or (len(self.t1) > self.p)):
+                old = self.t1.pop()
+                self.b1.appendleft(old)
+            else:
+                if self.t2:
+                    old = self.t2.pop()
+                    self.b2.appendleft(old)
+                else:
+                    old = self.t1.pop()
+                    self.b1.appendleft(old)
+            self.used_size -= self.table[old].value_size
+            del self.table[old]
+
+    def _lookup(self, trace_record, key, hash):
+        # Case I: key is in T1 or T2.
+        #   Move key to MRU position in T2.
+        if key in self.t1:
+            self.t1.remove(key)
+            self.t2.appendleft(key)
+            return True
+
+        if key in self.t2:
+            self.t2.remove(key)
+            self.t2.appendleft(key)
+            return True
+        return False
+
+    def _evict(self, trace_record, key, hash, value_size):
+        # Case II: key is in B1
+        #   Move x from B1 to the MRU position in T2 (also fetch x to the cache).
+        if key in self.b1:
+            self.p = min(self.c, self.p + max(len(self.b2) / len(self.b1), 1))
+            self._replace(key, value_size)
+            self.b1.remove(key)
+            self.t2.appendleft(key)
+            return
+
+        # Case III: key is in B2
+        #   Move x from B2 to the MRU position in T2 (also fetch x to the cache).
+        if key in self.b2:
+            self.p = max(0, self.p - max(len(self.b1) / len(self.b2), 1))
+            self._replace(key, value_size)
+            self.b2.remove(key)
+            self.t2.appendleft(key)
+            return
+
+        # Case IV: key is not in (T1 u B1 u T2 u B2)
+        self._replace(key, value_size)
+        while len(self.t1) + len(self.b1) >= self.c and self.b1:
+            self.b1.pop()
+
+        total = len(self.t1) + len(self.b1) + len(self.t2) + len(self.b2)
+        while total >= (2 * self.c) and self.b2:
+            self.b2.pop()
+            total -= 1
+        # Finally, move it to MRU position in T1.
+        self.t1.appendleft(key)
+        return
+
+    def _insert(self, trace_record, key, hash, value_size):
+        self.table[key] = CacheEntry(
+            value_size,
+            trace_record.cf_id,
+            trace_record.level,
+            trace_record.block_type,
+            trace_record.table_id,
+            0,
+            trace_record.access_time,
+        )
+
+    def _should_admit(self, trace_record, key, hash, value_size):
+        return True
+
+    def cache_name(self):
+        if self.enable_cache_row_key:
+            return "Hybrid Adaptive Replacement Cache (arc_hybrid)"
+        return "Adaptive Replacement Cache (arc)"
+
+
+class LRUCache(Cache):
+    """
+    A strict LRU queue.
+    """
+
+    def __init__(self, cache_size, enable_cache_row_key):
+        super(LRUCache, self).__init__(cache_size, enable_cache_row_key)
+        self.table = {}
+        self.lru = Deque()
+
+    def cache_name(self):
+        if self.enable_cache_row_key:
+            return "Hybrid LRU (lru_hybrid)"
+        return "LRU (lru)"
+
+    def _lookup(self, trace_record, key, hash):
+        if key not in self.table:
+            return False
+        # A cache hit. Update LRU queue.
+        self.lru.remove(key)
+        self.lru.appendleft(key)
+        return True
+
+    def _evict(self, trace_record, key, hash, value_size):
+        while self.used_size + value_size > self.cache_size:
+            evict_key = self.lru.pop()
+            self.used_size -= self.table[evict_key].value_size
+            del self.table[evict_key]
+
+    def _insert(self, trace_record, key, hash, value_size):
+        self.table[key] = CacheEntry(
+            value_size,
+            trace_record.cf_id,
+            trace_record.level,
+            trace_record.block_type,
+            trace_record.table_id,
+            0,
+            trace_record.access_time,
+        )
+        self.lru.appendleft(key)
+
+    def _should_admit(self, trace_record, key, hash, value_size):
+        return True
+
+
+class TraceCache(Cache):
+    """
+    A trace cache. Lookup returns true if the trace observes a cache hit.
+    It is used to maintain cache hits observed in the trace.
+    """
+
+    def __init__(self, cache_size):
+        super(TraceCache, self).__init__(cache_size, enable_cache_row_key=0)
+
+    def _lookup(self, trace_record, key, hash):
+        return trace_record.is_hit
+
+    def _evict(self, trace_record, key, hash, value_size):
+        pass
+
+    def _insert(self, trace_record, key, hash, value_size):
+        pass
+
+    def _should_admit(self, trace_record, key, hash, value_size):
+        return False
+
+    def cache_name(self):
+        return "Trace"
+
+
+def parse_cache_size(cs):
+    cs = cs.replace("\n", "")
+    if cs[-1] == "M":
+        return int(cs[: len(cs) - 1]) * 1024 * 1024
+    if cs[-1] == "G":
+        return int(cs[: len(cs) - 1]) * 1024 * 1024 * 1024
+    if cs[-1] == "T":
+        return int(cs[: len(cs) - 1]) * 1024 * 1024 * 1024 * 1024
+    return int(cs)
+
+
+def create_cache(cache_type, cache_size, downsample_size):
+    cache_size = cache_size / downsample_size
+    enable_cache_row_key = 0
+    if "hybridn" in cache_type:
+        enable_cache_row_key = 2
+        cache_type = cache_type[:-8]
+    if "hybrid" in cache_type:
+        enable_cache_row_key = 1
+        cache_type = cache_type[:-7]
+    if cache_type == "ts":
+        return ThompsonSamplingCache(
+            cache_size,
+            enable_cache_row_key,
+            [LRUPolicy(), LFUPolicy(), HyperbolicPolicy()],
+            cost_class_label=None,
+        )
+    elif cache_type == "linucb":
+        return LinUCBCache(
+            cache_size,
+            enable_cache_row_key,
+            [LRUPolicy(), LFUPolicy(), HyperbolicPolicy()],
+            cost_class_label=None,
+        )
+    elif cache_type == "pylru":
+        return ThompsonSamplingCache(
+            cache_size, enable_cache_row_key, [LRUPolicy()], cost_class_label=None
+        )
+    elif cache_type == "pymru":
+        return ThompsonSamplingCache(
+            cache_size, enable_cache_row_key, [MRUPolicy()], cost_class_label=None
+        )
+    elif cache_type == "pylfu":
+        return ThompsonSamplingCache(
+            cache_size, enable_cache_row_key, [LFUPolicy()], cost_class_label=None
+        )
+    elif cache_type == "pyhb":
+        return ThompsonSamplingCache(
+            cache_size,
+            enable_cache_row_key,
+            [HyperbolicPolicy()],
+            cost_class_label=None,
+        )
+    elif cache_type == "pycctbbt":
+        return ThompsonSamplingCache(
+            cache_size,
+            enable_cache_row_key,
+            [CostClassPolicy()],
+            cost_class_label="table_bt",
+        )
+    elif cache_type == "pycccf":
+        return ThompsonSamplingCache(
+            cache_size, enable_cache_row_key, [CostClassPolicy()], cost_class_label="cf"
+        )
+    elif cache_type == "pycctblevelbt":
+        return ThompsonSamplingCache(
+            cache_size,
+            enable_cache_row_key,
+            [CostClassPolicy()],
+            cost_class_label="table_level_bt",
+        )
+    elif cache_type == "pycccfbt":
+        return ThompsonSamplingCache(
+            cache_size,
+            enable_cache_row_key,
+            [CostClassPolicy()],
+            cost_class_label="cf_bt",
+        )
+    elif cache_type == "pycctb":
+        return ThompsonSamplingCache(
+            cache_size,
+            enable_cache_row_key,
+            [CostClassPolicy()],
+            cost_class_label="table",
+        )
+    elif cache_type == "pyccbt":
+        return ThompsonSamplingCache(
+            cache_size, enable_cache_row_key, [CostClassPolicy()], cost_class_label="bt"
+        )
+    elif cache_type == "opt":
+        if enable_cache_row_key:
+            print("opt does not support hybrid mode.")
+            assert False
+        return OPTCache(cache_size)
+    elif cache_type == "trace":
+        if enable_cache_row_key:
+            print("trace does not support hybrid mode.")
+            assert False
+        return TraceCache(cache_size)
+    elif cache_type == "lru":
+        return LRUCache(cache_size, enable_cache_row_key)
+    elif cache_type == "arc":
+        return ARCCache(cache_size, enable_cache_row_key)
+    elif cache_type == "gdsize":
+        return GDSizeCache(cache_size, enable_cache_row_key)
+    else:
+        print("Unknown cache type {}".format(cache_type))
+        assert False
+    return None
+
+
+class BlockAccessTimeline:
+    """
+    BlockAccessTimeline stores all accesses of a block.
+    """
+
+    def __init__(self):
+        self.accesses = []
+        self.current_access_index = 1
+
+    def get_next_access(self):
+        if self.current_access_index == len(self.accesses):
+            return sys.maxsize
+        next_access_seq_no = self.accesses[self.current_access_index]
+        self.current_access_index += 1
+        return next_access_seq_no
+
+
+def percent(e1, e2):
+    if e2 == 0:
+        return -1
+    return float(e1) * 100.0 / float(e2)
+
+
+def is_target_cf(access_cf, target_cf_name):
+    if target_cf_name == "all":
+        return True
+    return access_cf == target_cf_name
+
+
+def run(
+    trace_file_path,
+    cache_type,
+    cache,
+    warmup_seconds,
+    max_accesses_to_process,
+    target_cf_name,
+):
+    warmup_complete = False
+    trace_miss_ratio_stats = MissRatioStats(kSecondsInMinute)
+    access_seq_no = 0
+    time_interval = 1
+    start_time = time.time()
+    trace_start_time = 0
+    trace_duration = 0
+    is_opt_cache = False
+    if cache.cache_name() == "Belady MIN (opt)":
+        is_opt_cache = True
+
+    block_access_timelines = {}
+    num_no_inserts = 0
+    num_blocks_with_no_size = 0
+    num_inserts_block_with_no_size = 0
+
+    if is_opt_cache:
+        # Read all blocks in memory and stores their access times so that OPT
+        # can use this information to evict the cached key which next access is
+        # the furthest in the future.
+        print("Preprocessing block traces.")
+        with open(trace_file_path, "r") as trace_file:
+            for line in trace_file:
+                if (
+                    max_accesses_to_process != -1
+                    and access_seq_no > max_accesses_to_process
+                ):
+                    break
+                ts = line.split(",")
+                timestamp = int(ts[0])
+                cf_name = ts[5]
+                if not is_target_cf(cf_name, target_cf_name):
+                    continue
+                if trace_start_time == 0:
+                    trace_start_time = timestamp
+                trace_duration = timestamp - trace_start_time
+                block_id = int(ts[1])
+                block_size = int(ts[3])
+                no_insert = int(ts[9])
+                if block_id not in block_access_timelines:
+                    block_access_timelines[block_id] = BlockAccessTimeline()
+                    if block_size == 0:
+                        num_blocks_with_no_size += 1
+                block_access_timelines[block_id].accesses.append(access_seq_no)
+                access_seq_no += 1
+                if no_insert == 1:
+                    num_no_inserts += 1
+                if no_insert == 0 and block_size == 0:
+                    num_inserts_block_with_no_size += 1
+                if access_seq_no % 100 != 0:
+                    continue
+                now = time.time()
+                if now - start_time > time_interval * 10:
+                    print(
+                        "Take {} seconds to process {} trace records with trace "
+                        "duration of {} seconds. Throughput: {} records/second.".format(
+                            now - start_time,
+                            access_seq_no,
+                            trace_duration / 1000000,
+                            access_seq_no / (now - start_time),
+                        )
+                    )
+                    time_interval += 1
+            print(
+                "Trace contains {0} blocks, {1}({2:.2f}%) blocks with no size."
+                "{3} accesses, {4}({5:.2f}%) accesses with no_insert,"
+                "{6}({7:.2f}%) accesses that want to insert but block size is 0.".format(
+                    len(block_access_timelines),
+                    num_blocks_with_no_size,
+                    percent(num_blocks_with_no_size, len(block_access_timelines)),
+                    access_seq_no,
+                    num_no_inserts,
+                    percent(num_no_inserts, access_seq_no),
+                    num_inserts_block_with_no_size,
+                    percent(num_inserts_block_with_no_size, access_seq_no),
+                )
+            )
+
+    access_seq_no = 0
+    time_interval = 1
+    start_time = time.time()
+    trace_start_time = 0
+    trace_duration = 0
+    print("Running simulated {} cache on block traces.".format(cache.cache_name()))
+    with open(trace_file_path, "r") as trace_file:
+        for line in trace_file:
+            if (
+                max_accesses_to_process != -1
+                and access_seq_no > max_accesses_to_process
+            ):
+                break
+            if access_seq_no % 1000000 == 0:
+                # Force a python gc periodically to reduce memory usage.
+                gc.collect()
+            ts = line.split(",")
+            timestamp = int(ts[0])
+            cf_name = ts[5]
+            if not is_target_cf(cf_name, target_cf_name):
+                continue
+            if trace_start_time == 0:
+                trace_start_time = timestamp
+            trace_duration = timestamp - trace_start_time
+            if (
+                not warmup_complete
+                and warmup_seconds > 0
+                and trace_duration > warmup_seconds * 1000000
+            ):
+                cache.miss_ratio_stats.reset_counter()
+                warmup_complete = True
+            next_access_seq_no = 0
+            block_id = int(ts[1])
+            if is_opt_cache:
+                next_access_seq_no = block_access_timelines[block_id].get_next_access()
+            record = TraceRecord(
+                access_time=int(ts[0]),
+                block_id=int(ts[1]),
+                block_type=int(ts[2]),
+                block_size=int(ts[3]),
+                cf_id=int(ts[4]),
+                cf_name=ts[5],
+                level=int(ts[6]),
+                fd=int(ts[7]),
+                caller=int(ts[8]),
+                no_insert=int(ts[9]),
+                get_id=int(ts[10]),
+                key_id=int(ts[11]),
+                kv_size=int(ts[12]),
+                is_hit=int(ts[13]),
+                referenced_key_exist_in_block=int(ts[14]),
+                num_keys_in_block=int(ts[15]),
+                table_id=int(ts[16]),
+                seq_number=int(ts[17]),
+                block_key_size=int(ts[18]),
+                key_size=int(ts[19]),
+                block_offset_in_file=int(ts[20]),
+                next_access_seq_no=next_access_seq_no,
+            )
+            trace_miss_ratio_stats.update_metrics(
+                record.access_time, is_hit=record.is_hit, miss_bytes=record.block_size
+            )
+            cache.access(record)
+            access_seq_no += 1
+            del record
+            del ts
+            if access_seq_no % 100 != 0:
+                continue
+            # Report progress every 10 seconds.
+            now = time.time()
+            if now - start_time > time_interval * 10:
+                print(
+                    "Take {} seconds to process {} trace records with trace "
+                    "duration of {} seconds. Throughput: {} records/second. "
+                    "Trace miss ratio {}".format(
+                        now - start_time,
+                        access_seq_no,
+                        trace_duration / 1000000,
+                        access_seq_no / (now - start_time),
+                        trace_miss_ratio_stats.miss_ratio(),
+                    )
+                )
+                time_interval += 1
+                print(
+                    "{},0,0,{},{},{}".format(
+                        cache_type,
+                        cache.cache_size,
+                        cache.miss_ratio_stats.miss_ratio(),
+                        cache.miss_ratio_stats.num_accesses,
+                    )
+                )
+    now = time.time()
+    print(
+        "Take {} seconds to process {} trace records with trace duration of {} "
+        "seconds. Throughput: {} records/second. Trace miss ratio {}".format(
+            now - start_time,
+            access_seq_no,
+            trace_duration / 1000000,
+            access_seq_no / (now - start_time),
+            trace_miss_ratio_stats.miss_ratio(),
+        )
+    )
+    print(
+        "{},0,0,{},{},{}".format(
+            cache_type,
+            cache.cache_size,
+            cache.miss_ratio_stats.miss_ratio(),
+            cache.miss_ratio_stats.num_accesses,
+        )
+    )
+    return trace_start_time, trace_duration
+
+
+def report_stats(
+    cache,
+    cache_type,
+    cache_size,
+    target_cf_name,
+    result_dir,
+    trace_start_time,
+    trace_end_time,
+):
+    cache_label = "{}-{}-{}".format(cache_type, cache_size, target_cf_name)
+    with open("{}/data-ml-mrc-{}".format(result_dir, cache_label), "w+") as mrc_file:
+        mrc_file.write(
+            "{},0,0,{},{},{}\n".format(
+                cache_type,
+                cache_size,
+                cache.miss_ratio_stats.miss_ratio(),
+                cache.miss_ratio_stats.num_accesses,
+            )
+        )
+
+    cache_stats = [
+        cache.per_second_miss_ratio_stats,
+        cache.miss_ratio_stats,
+        cache.per_hour_miss_ratio_stats,
+    ]
+    for i in range(len(cache_stats)):
+        avg_miss_bytes, p95_miss_bytes = cache_stats[i].compute_miss_bytes()
+
+        with open(
+            "{}/data-ml-avgmb-{}-{}".format(
+                result_dir, cache_stats[i].time_unit, cache_label
+            ),
+            "w+",
+        ) as mb_file:
+            mb_file.write(
+                "{},0,0,{},{}\n".format(cache_type, cache_size, avg_miss_bytes)
+            )
+
+        with open(
+            "{}/data-ml-p95mb-{}-{}".format(
+                result_dir, cache_stats[i].time_unit, cache_label
+            ),
+            "w+",
+        ) as mb_file:
+            mb_file.write(
+                "{},0,0,{},{}\n".format(cache_type, cache_size, p95_miss_bytes)
+            )
+
+        cache_stats[i].write_miss_timeline(
+            cache_type,
+            cache_size,
+            target_cf_name,
+            result_dir,
+            trace_start_time,
+            trace_end_time,
+        )
+        cache_stats[i].write_miss_ratio_timeline(
+            cache_type,
+            cache_size,
+            target_cf_name,
+            result_dir,
+            trace_start_time,
+            trace_end_time,
+        )
+
+    if not cache.is_ml_cache():
+        return
+
+    policy_stats = [cache.policy_stats, cache.per_hour_policy_stats]
+    for i in range(len(policy_stats)):
+        policy_stats[i].write_policy_timeline(
+            cache_type,
+            cache_size,
+            target_cf_name,
+            result_dir,
+            trace_start_time,
+            trace_end_time,
+        )
+        policy_stats[i].write_policy_ratio_timeline(
+            cache_type,
+            cache_size,
+            target_cf_name,
+            result_dir,
+            trace_start_time,
+            trace_end_time,
+        )
+
+
+if __name__ == "__main__":
+    if len(sys.argv) <= 8:
+        print(
+            "Must provide 8 arguments.\n"
+            "1) Cache type (ts, linucb, arc, lru, opt, pylru, pymru, pylfu, "
+            "pyhb, gdsize, trace). One may evaluate the hybrid row_block cache "
+            "by appending '_hybrid' to a cache_type, e.g., ts_hybrid. "
+            "Note that hybrid is not supported with opt and trace. \n"
+            "2) Cache size (xM, xG, xT).\n"
+            "3) The sampling frequency used to collect the trace. (The "
+            "simulation scales down the cache size by the sampling frequency).\n"
+            "4) Warmup seconds (The number of seconds used for warmup).\n"
+            "5) Trace file path.\n"
+            "6) Result directory (A directory that saves generated results)\n"
+            "7) Max number of accesses to process\n"
+            "8) The target column family. (The simulation will only run "
+            "accesses on the target column family. If it is set to all, "
+            "it will run against all accesses.)"
+        )
+        exit(1)
+    print("Arguments: {}".format(sys.argv))
+    cache_type = sys.argv[1]
+    cache_size = parse_cache_size(sys.argv[2])
+    downsample_size = int(sys.argv[3])
+    warmup_seconds = int(sys.argv[4])
+    trace_file_path = sys.argv[5]
+    result_dir = sys.argv[6]
+    max_accesses_to_process = int(sys.argv[7])
+    target_cf_name = sys.argv[8]
+    cache = create_cache(cache_type, cache_size, downsample_size)
+    trace_start_time, trace_duration = run(
+        trace_file_path,
+        cache_type,
+        cache,
+        warmup_seconds,
+        max_accesses_to_process,
+        target_cf_name,
+    )
+    trace_end_time = trace_start_time + trace_duration
+    report_stats(
+        cache,
+        cache_type,
+        cache_size,
+        target_cf_name,
+        result_dir,
+        trace_start_time,
+        trace_end_time,
+    )