source/configs/clickhouse-config.xml

<?xml version="1.0"?>
<!--
  ClickHouse Server Configuration
  =================================
  Hardware: AMD EPYC 9365 36-Core (72 threads), 2.2 TB RAM, 2x NUMA nodes
  Purpose : Data analytics for LLM training pipeline + factory sensor data (CRM project)
  Generated: 2026-03-01
-->
<clickhouse>

    <!-- =========================================================
         Network / Listen
         Listen on localhost only — GPU training node is local only
    ========================================================= -->
    <listen_host>127.0.0.1</listen_host>
    <http_port>8123</http_port>
    <tcp_port>9000</tcp_port>
    <interserver_http_port>9009</interserver_http_port>

    <!-- =========================================================
         Paths
         Data on GPFS (20 TB, 18 TB free) for large datasets.
         Tmp / logs on local /tmp to reduce GPFS small-file pressure.
    ========================================================= -->
    <path>/PROJECT/0325120031_A/ghong/taketimes/clickhouse-data/</path>
    <tmp_path>/tmp/clickhouse-tmp/</tmp_path>

    <!-- =========================================================
         Logging
    ========================================================= -->
    <logger>
        <level>information</level>
        <log>/tmp/clickhouse/logs/clickhouse-server.log</log>
        <errorlog>/tmp/clickhouse/logs/clickhouse-server.err.log</errorlog>
        <!-- Rotate at 512 MB, keep 10 files -->
        <size>536870912</size>
        <count>10</count>
    </logger>

    <!-- =========================================================
         Memory — server-level cap for all queries combined
    ========================================================= -->
    <max_server_memory_usage>536870912000</max_server_memory_usage>

    <!-- =========================================================
         CPU / Thread Pools (server-level settings)
         Physical cores: 36 per socket, 72 total (2 NUMA nodes).
         Reserve half for GPU training → 36 threads for ClickHouse.
    ========================================================= -->
    <background_pool_size>18</background_pool_size>
    <background_merges_mutations_concurrency_ratio>2</background_merges_mutations_concurrency_ratio>
    <background_move_pool_size>4</background_move_pool_size>
    <background_fetches_pool_size>4</background_fetches_pool_size>
    <background_schedule_pool_size>8</background_schedule_pool_size>
    <background_common_pool_size>8</background_common_pool_size>

    <!-- =========================================================
         Concurrency
    ========================================================= -->
    <max_concurrent_queries>100</max_concurrent_queries>
    <max_waiting_queries>50</max_waiting_queries>

    <!-- =========================================================
         Caches
         mark_cache    : 10 GB  — indexes for MergeTree parts
         uncompressed  : 20 GB  — decompressed block cache
         query_cache   : 2  GB  — optional query result cache
         NUMA note: ClickHouse allocates via jemalloc with NUMA
           awareness; no extra config needed beyond thread binding.
    ========================================================= -->
    <mark_cache_size>10737418240</mark_cache_size>
    <uncompressed_cache_size>21474836480</uncompressed_cache_size>
    <!-- Available from ClickHouse 23.x -->
    <query_cache>
        <max_size_in_bytes>2147483648</max_size_in_bytes>
        <max_entries>1024</max_entries>
        <max_entry_size_in_bytes>104857600</max_entry_size_in_bytes>
        <max_entry_size_in_rows>30000000</max_entry_size_in_rows>
    </query_cache>
    <!-- Compiled expression cache -->
    <compiled_expression_cache_size>134217728</compiled_expression_cache_size>
    <compiled_expression_cache_elements_size>10000</compiled_expression_cache_elements_size>

    <!-- =========================================================
         I/O
         GPFS is a parallel filesystem — large sequential reads are
         efficient; use aggressive read-ahead and prefetch.
         NVMe local disks can be used for tmp / intermediate data.
    ========================================================= -->
    <!-- MergeTree concurrent read settings moved to <profiles><default> -->
    <!-- Async reads from object/POSIX storage -->
    <asynchronous_metrics_update_period_s>60</asynchronous_metrics_update_period_s>

    <!-- Async Insert settings moved to <profiles><default> below -->

    <!-- =========================================================
         MergeTree Storage Settings
    ========================================================= -->
    <merge_tree>
        <!-- Bytes: prefer larger parts on GPFS to reduce metadata overhead -->
        <max_bytes_to_merge_at_max_space_in_pool>161061273600</max_bytes_to_merge_at_max_space_in_pool>
        <!-- Allow up to 300 parts per partition before slowing inserts -->
        <parts_to_throw_insert>300</parts_to_throw_insert>
        <parts_to_delay_insert>150</parts_to_delay_insert>
        <!-- Use AVX-512 SIMD for sorting / hashing where available -->
        <use_minimalistic_part_header_in_zookeeper>1</use_minimalistic_part_header_in_zookeeper>
        <!-- Keep deleted data up to 8 hours before final cleanup -->
        <old_parts_lifetime>28800</old_parts_lifetime>
    </merge_tree>

    <!-- =========================================================
         Compression
         LZ4 default (fast, AVX2/AVX-512 accelerated).
         ZSTD level 3 for cold / archival tables — trade CPU for space.
    ========================================================= -->
    <compression>
        <!-- Hot data: LZ4 -->
        <case>
            <min_part_size>1073741824</min_part_size>
            <min_part_size_ratio>0.01</min_part_size_ratio>
            <method>lz4</method>
        </case>
        <!-- Very large parts: ZSTD for better ratio -->
        <case>
            <min_part_size>10737418240</min_part_size>
            <min_part_size_ratio>0.1</min_part_size_ratio>
            <method>zstd</method>
            <level>3</level>
        </case>
    </compression>

    <!-- =========================================================
         Users / Access Control
         Single local user, no password (localhost-only listen).
         See users.xml (or inline below) for quota/profile.
    ========================================================= -->
    <users>
        <default>
            <password></password>
            <networks>
                <ip>127.0.0.1/8</ip>
                <ip>::1</ip>
            </networks>
            <profile>default</profile>
            <quota>default</quota>
            <!-- Allow DDL from default user -->
            <access_management>1</access_management>
        </default>
    </users>

    <profiles>
        <default>
            <!-- Memory per query: 500 GB -->
            <max_memory_usage>536870912000</max_memory_usage>
            <max_bytes_before_external_group_by>483183820800</max_bytes_before_external_group_by>
            <max_bytes_before_external_sort>483183820800</max_bytes_before_external_sort>
            <!-- Threads per query: half of 72 cores -->
            <max_threads>36</max_threads>
            <max_concurrent_queries_for_user>100</max_concurrent_queries_for_user>
            <use_uncompressed_cache>1</use_uncompressed_cache>
            <!-- O_DIRECT for large scans -->
            <min_bytes_to_use_direct_io>10737418240</min_bytes_to_use_direct_io>
            <!-- SIMD JSON parsing (AVX-512) -->
            <input_format_parallel_parsing>1</input_format_parallel_parsing>
            <output_format_parallel_formatting>1</output_format_parallel_formatting>
            <!-- Async insert for sensor/CRM streaming -->
            <async_insert>1</async_insert>
            <async_insert_max_data_size>33554432</async_insert_max_data_size>
            <async_insert_busy_timeout_ms>200</async_insert_busy_timeout_ms>
            <async_insert_deduplicate>0</async_insert_deduplicate>
            <wait_for_async_insert>1</wait_for_async_insert>
            <wait_for_async_insert_timeout>5</wait_for_async_insert_timeout>
            <!-- MergeTree concurrent read -->
            <merge_tree_min_rows_for_concurrent_read>20000</merge_tree_min_rows_for_concurrent_read>
            <merge_tree_min_bytes_for_concurrent_read>24117248</merge_tree_min_bytes_for_concurrent_read>
        </default>
    </profiles>

    <quotas>
        <default>
            <interval>
                <duration>3600</duration>
                <queries>0</queries>
                <errors>0</errors>
                <result_rows>0</result_rows>
                <read_rows>0</read_rows>
                <execution_time>0</execution_time>
            </interval>
        </default>
    </quotas>

    <!-- =========================================================
         Distributed DDL (single-node — disable ZooKeeper dependency)
    ========================================================= -->
    <!-- No ZooKeeper configured; replicated tables use ReplicatedMergeTree
         only if ZK is added later. Commenting out to avoid startup warnings.
    <zookeeper>
        <node>
            <host>localhost</host>
            <port>2181</port>
        </node>
    </zookeeper>
    -->

    <!-- =========================================================
         Timezone
    ========================================================= -->
    <timezone>Asia/Seoul</timezone>

    <!-- =========================================================
         Query Log / System Tables
         Keep 30 days of query history for pipeline debugging.
    ========================================================= -->
    <query_log>
        <database>system</database>
        <table>query_log</table>
        <partition_by>toYYYYMM(event_date)</partition_by>
        <ttl>event_date + INTERVAL 30 DAY</ttl>
        <flush_interval_milliseconds>7500</flush_interval_milliseconds>
        <max_size_rows>1048576</max_size_rows>
    </query_log>

    <query_thread_log>
        <database>system</database>
        <table>query_thread_log</table>
        <partition_by>toYYYYMM(event_date)</partition_by>
        <ttl>event_date + INTERVAL 7 DAY</ttl>
        <flush_interval_milliseconds>7500</flush_interval_milliseconds>
    </query_thread_log>

    <part_log>
        <database>system</database>
        <table>part_log</table>
        <partition_by>toYYYYMM(event_date)</partition_by>
        <ttl>event_date + INTERVAL 14 DAY</ttl>
        <flush_interval_milliseconds>5000</flush_interval_milliseconds>
    </part_log>

    <trace_log>
        <database>system</database>
        <table>trace_log</table>
        <partition_by>toYYYYMM(event_date)</partition_by>
        <ttl>event_date + INTERVAL 7 DAY</ttl>
        <flush_interval_milliseconds>7500</flush_interval_milliseconds>
    </trace_log>

    <metric_log>
        <database>system</database>
        <table>metric_log</table>
        <flush_interval_milliseconds>7500</flush_interval_milliseconds>
        <collect_interval_milliseconds>1000</collect_interval_milliseconds>
        <ttl>event_date + INTERVAL 7 DAY</ttl>
    </metric_log>

    <asynchronous_metric_log>
        <database>system</database>
        <table>asynchronous_metric_log</table>
        <flush_interval_milliseconds>7500</flush_interval_milliseconds>
        <ttl>event_date + INTERVAL 7 DAY</ttl>
    </asynchronous_metric_log>

    <!-- =========================================================
         Crash Handler
    ========================================================= -->
    <core_dump>
        <size_limit>0</size_limit>
    </core_dump>

    <!-- =========================================================
         Keeper (built-in, single-node mode — replaces ZooKeeper
         if you want ReplicatedMergeTree without external ZK).
         Uncomment if needed.
    ========================================================= -->
    <!--
    <keeper_server>
        <tcp_port>9181</tcp_port>
        <server_id>1</server_id>
        <log_storage_path>/PROJECT/0325120031_A/ghong/taketimes/clickhouse-data/keeper/logs</log_storage_path>
        <snapshot_storage_path>/PROJECT/0325120031_A/ghong/taketimes/clickhouse-data/keeper/snapshots</snapshot_storage_path>
        <coordination_settings>
            <operation_timeout_ms>10000</operation_timeout_ms>
            <session_timeout_ms>30000</session_timeout_ms>
            <raft_logs_level>warning</raft_logs_level>
        </coordination_settings>
        <raft_configuration>
            <server>
                <id>1</id>
                <hostname>localhost</hostname>
                <port>9444</port>
            </server>
        </raft_configuration>
    </keeper_server>
    -->

    <!-- =========================================================
         AVX-512 / SIMD hints
         ClickHouse auto-detects CPUID at runtime; these flags are
         informational comments — no XML knobs needed.
         Detected: avx512f, avx512bw, avx512vl, avx512_vnni, avx512_bf16
         Used in: LZ4 compression, hash aggregation, sorting, filters.
    ========================================================= -->

    <!-- =========================================================
         Miscellaneous
    ========================================================= -->
    <!-- Skip strict settings check for forward-compat -->
    <skip_check_for_incorrect_settings>1</skip_check_for_incorrect_settings>
    <!-- Graceful shutdown: wait up to 60 s for running queries -->
    <shutdown_wait_unfinished>60</shutdown_wait_unfinished>
    <!-- Send anonymous usage statistics: off for private server -->
    <send_crash_reports>
        <enabled>false</enabled>
    </send_crash_reports>

</clickhouse>