path: root/src/latency.c

/* The latency monitor allows to easily observe the sources of latency
 * in a Redis instance using the LATENCY command. Different latency
 * sources are monitored, like disk I/O, execution of commands, fork
 * system call, and so forth.
 *
 * ----------------------------------------------------------------------------
 *
 * Copyright (c) 2014, Salvatore Sanfilippo <antirez at gmail dot com>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 *   * Redistributions of source code must retain the above copyright notice,
 *     this list of conditions and the following disclaimer.
 *   * Redistributions in binary form must reproduce the above copyright
 *     notice, this list of conditions and the following disclaimer in the
 *     documentation and/or other materials provided with the distribution.
 *   * Neither the name of Redis nor the names of its contributors may be used
 *     to endorse or promote products derived from this software without
 *     specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 */

#include "server.h"

/* Dictionary type for latency events. */
int dictStringKeyCompare(dict *d, const void *key1, const void *key2) {
    UNUSED(d);
    return strcmp(key1,key2) == 0;
}

uint64_t dictStringHash(const void *key) {
    return dictGenHashFunction(key, strlen(key));
}

void dictVanillaFree(dict *d, void *val);

dictType latencyTimeSeriesDictType = {
    dictStringHash,             /* hash function */
    NULL,                       /* key dup */
    NULL,                       /* val dup */
    dictStringKeyCompare,       /* key compare */
    dictVanillaFree,            /* key destructor */
    dictVanillaFree,            /* val destructor */
    NULL                        /* allow to expand */
};

/* ------------------------- Utility functions ------------------------------ */

#ifdef __linux__
#include <sys/prctl.h>
/* Returns 1 if Transparent Huge Pages support is enabled in the kernel.
 * Otherwise (or if we are unable to check) 0 is returned. */
int THPIsEnabled(void) {
    char buf[1024];

    FILE *fp = fopen("/sys/kernel/mm/transparent_hugepage/enabled","r");
    if (!fp) return 0;
    if (fgets(buf,sizeof(buf),fp) == NULL) {
        fclose(fp);
        return 0;
    }
    fclose(fp);
    return (strstr(buf,"[always]") != NULL) ? 1 : 0;
}

/* since linux-3.5, kernel supports to set the state of the "THP disable" flag
 * for the calling thread. PR_SET_THP_DISABLE is defined in linux/prctl.h */
int THPDisable(void) {
    int ret = -EINVAL;

    if (!server.disable_thp)
        return ret;

#ifdef PR_SET_THP_DISABLE
    ret = prctl(PR_SET_THP_DISABLE, 1, 0, 0, 0);
#endif

    return ret;
}
#endif

/* Report the amount of AnonHugePages in smap, in bytes. If the return
 * value of the function is non-zero, the process is being targeted by
 * THP support, and is likely to have memory usage / latency issues. */
int THPGetAnonHugePagesSize(void) {
    return zmalloc_get_smap_bytes_by_field("AnonHugePages:",-1);
}

/* ---------------------------- Latency API --------------------------------- */

/* Latency monitor initialization. We just need to create the dictionary
 * of time series, each time series is created on demand in order to avoid
 * having a fixed list to maintain. */
void latencyMonitorInit(void) {
    server.latency_events = dictCreate(&latencyTimeSeriesDictType);
}

/* Add the specified sample to the specified time series "event".
 * This function is usually called via latencyAddSampleIfNeeded(), that
 * is a macro that only adds the sample if the latency is higher than
 * server.latency_monitor_threshold. */
void latencyAddSample(const char *event, mstime_t latency) {
    struct latencyTimeSeries *ts = dictFetchValue(server.latency_events,event);
    time_t now = time(NULL);
    int prev;

    /* Create the time series if it does not exist. */
    if (ts == NULL) {
        ts = zmalloc(sizeof(*ts));
        ts->idx = 0;
        ts->max = 0;
        memset(ts->samples,0,sizeof(ts->samples));
        dictAdd(server.latency_events,zstrdup(event),ts);
    }

    if (latency > ts->max) ts->max = latency;

    /* If the previous sample is in the same second, we update our old sample
     * if this latency is > of the old one, or just return. */
    prev = (ts->idx + LATENCY_TS_LEN - 1) % LATENCY_TS_LEN;
    if (ts->samples[prev].time == now) {
        if (latency > ts->samples[prev].latency)
            ts->samples[prev].latency = latency;
        return;
    }

    ts->samples[ts->idx].time = now;
    ts->samples[ts->idx].latency = latency;

    ts->idx++;
    if (ts->idx == LATENCY_TS_LEN) ts->idx = 0;
}

/* Reset data for the specified event, or all the events data if 'event' is
 * NULL.
 *
 * Note: this is O(N) even when event_to_reset is not NULL because makes
 * the code simpler and we have a small fixed max number of events. */
int latencyResetEvent(char *event_to_reset) {
    dictIterator *di;
    dictEntry *de;
    int resets = 0;

    di = dictGetSafeIterator(server.latency_events);
    while((de = dictNext(di)) != NULL) {
        char *event = dictGetKey(de);

        if (event_to_reset == NULL || strcasecmp(event,event_to_reset) == 0) {
            dictDelete(server.latency_events, event);
            resets++;
        }
    }
    dictReleaseIterator(di);
    return resets;
}

/* ------------------------ Latency reporting (doctor) ---------------------- */

/* Analyze the samples available for a given event and return a structure
 * populate with different metrics, average, MAD, min, max, and so forth.
 * Check latency.h definition of struct latencyStats for more info.
 * If the specified event has no elements the structure is populate with
 * zero values. */
void analyzeLatencyForEvent(char *event, struct latencyStats *ls) {
    struct latencyTimeSeries *ts = dictFetchValue(server.latency_events,event);
    int j;
    uint64_t sum;

    ls->all_time_high = ts ? ts->max : 0;
    ls->avg = 0;
    ls->min = 0;
    ls->max = 0;
    ls->mad = 0;
    ls->samples = 0;
    ls->period = 0;
    if (!ts) return;

    /* First pass, populate everything but the MAD. */
    sum = 0;
    for (j = 0; j < LATENCY_TS_LEN; j++) {
        if (ts->samples[j].time == 0) continue;
        ls->samples++;
        if (ls->samples == 1) {
            ls->min = ls->max = ts->samples[j].latency;
        } else {
            if (ls->min > ts->samples[j].latency)
                ls->min = ts->samples[j].latency;
            if (ls->max < ts->samples[j].latency)
                ls->max = ts->samples[j].latency;
        }
        sum += ts->samples[j].latency;

        /* Track the oldest event time in ls->period. */
        if (ls->period == 0 || ts->samples[j].time < ls->period)
            ls->period = ts->samples[j].time;
    }

    /* So far avg is actually the sum of the latencies, and period is
     * the oldest event time. We need to make the first an average and
     * the second a range of seconds. */
    if (ls->samples) {
        ls->avg = sum / ls->samples;
        ls->period = time(NULL) - ls->period;
        if (ls->period == 0) ls->period = 1;
    }

    /* Second pass, compute MAD. */
    sum = 0;
    for (j = 0; j < LATENCY_TS_LEN; j++) {
        int64_t delta;

        if (ts->samples[j].time == 0) continue;
        delta = (int64_t)ls->avg - ts->samples[j].latency;
        if (delta < 0) delta = -delta;
        sum += delta;
    }
    if (ls->samples) ls->mad = sum / ls->samples;
}

/* Create a human readable report of latency events for this Redis instance. */
sds createLatencyReport(void) {
    sds report = sdsempty();
    int advise_better_vm = 0;       /* Better virtual machines. */
    int advise_slowlog_enabled = 0; /* Enable slowlog. */
    int advise_slowlog_tuning = 0;  /* Reconfigure slowlog. */
    int advise_slowlog_inspect = 0; /* Check your slowlog. */
    int advise_disk_contention = 0; /* Try to lower disk contention. */
    int advise_scheduler = 0;       /* Intrinsic latency. */
    int advise_data_writeback = 0;  /* data=writeback. */
    int advise_no_appendfsync = 0;  /* don't fsync during rewrites. */
    int advise_local_disk = 0;      /* Avoid remote disks. */
    int advise_ssd = 0;             /* Use an SSD drive. */
    int advise_write_load_info = 0; /* Print info about AOF and write load. */
    int advise_hz = 0;              /* Use higher HZ. */
    int advise_large_objects = 0;   /* Deletion of large objects. */
    int advise_mass_eviction = 0;   /* Avoid mass eviction of keys. */
    int advise_relax_fsync_policy = 0; /* appendfsync always is slow. */
    int advise_disable_thp = 0;     /* AnonHugePages detected. */
    int advices = 0;

    /* Return ASAP if the latency engine is disabled and it looks like it
     * was never enabled so far. */
    if (dictSize(server.latency_events) == 0 &&
        server.latency_monitor_threshold == 0)
    {
        report = sdscat(report,"I'm sorry, Dave, I can't do that. Latency monitoring is disabled in this Redis instance. You may use \"CONFIG SET latency-monitor-threshold <milliseconds>.\" in order to enable it. If we weren't in a deep space mission I'd suggest to take a look at https://redis.io/topics/latency-monitor.\n");
        return report;
    }

    /* Show all the events stats and add for each event some event-related
     * comment depending on the values. */
    dictIterator *di;
    dictEntry *de;
    int eventnum = 0;

    di = dictGetSafeIterator(server.latency_events);
    while((de = dictNext(di)) != NULL) {
        char *event = dictGetKey(de);
        struct latencyTimeSeries *ts = dictGetVal(de);
        struct latencyStats ls;

        if (ts == NULL) continue;
        eventnum++;
        if (eventnum == 1) {
            report = sdscat(report,"Dave, I have observed latency spikes in this Redis instance. You don't mind talking about it, do you Dave?\n\n");
        }
        analyzeLatencyForEvent(event,&ls);

        report = sdscatprintf(report,
            "%d. %s: %d latency spikes (average %lums, mean deviation %lums, period %.2f sec). Worst all time event %lums.",
            eventnum, event,
            ls.samples,
            (unsigned long) ls.avg,
            (unsigned long) ls.mad,
            (double) ls.period/ls.samples,
            (unsigned long) ts->max);

        /* Fork */
        if (!strcasecmp(event,"fork")) {
            char *fork_quality;
            if (server.stat_fork_rate < 10) {
                fork_quality = "terrible";
                advise_better_vm = 1;
                advices++;
            } else if (server.stat_fork_rate < 25) {
                fork_quality = "poor";
                advise_better_vm = 1;
                advices++;
            } else if (server.stat_fork_rate < 100) {
                fork_quality = "good";
            } else {
                fork_quality = "excellent";
            }
            report = sdscatprintf(report,
                " Fork rate is %.2f GB/sec (%s).", server.stat_fork_rate,
                fork_quality);
        }

        /* Potentially commands. */
        if (!strcasecmp(event,"command")) {
            if (server.slowlog_log_slower_than < 0) {
                advise_slowlog_enabled = 1;
                advices++;
            } else if (server.slowlog_log_slower_than/1000 >
                       server.latency_monitor_threshold)
            {
                advise_slowlog_tuning = 1;
                advices++;
            }
            advise_slowlog_inspect = 1;
            advise_large_objects = 1;
            advices += 2;
        }

        /* fast-command. */
        if (!strcasecmp(event,"fast-command")) {
            advise_scheduler = 1;
            advices++;
        }

        /* AOF and I/O. */
        if (!strcasecmp(event,"aof-write-pending-fsync")) {
            advise_local_disk = 1;
            advise_disk_contention = 1;
            advise_ssd = 1;
            advise_data_writeback = 1;
            advices += 4;
        }

        if (!strcasecmp(event,"aof-write-active-child")) {
            advise_no_appendfsync = 1;
            advise_data_writeback = 1;
            advise_ssd = 1;
            advices += 3;
        }

        if (!strcasecmp(event,"aof-write-alone")) {
            advise_local_disk = 1;
            advise_data_writeback = 1;
            advise_ssd = 1;
            advices += 3;
        }

        if (!strcasecmp(event,"aof-fsync-always")) {
            advise_relax_fsync_policy = 1;
            advices++;
        }

        if (!strcasecmp(event,"aof-fstat") ||
            !strcasecmp(event,"rdb-unlink-temp-file")) {
            advise_disk_contention = 1;
            advise_local_disk = 1;
            advices += 2;
        }

        if (!strcasecmp(event,"aof-rewrite-diff-write") ||
            !strcasecmp(event,"aof-rename")) {
            advise_write_load_info = 1;
            advise_data_writeback = 1;
            advise_ssd = 1;
            advise_local_disk = 1;
            advices += 4;
        }

        /* Expire cycle. */
        if (!strcasecmp(event,"expire-cycle")) {
            advise_hz = 1;
            advise_large_objects = 1;
            advices += 2;
        }

        /* Eviction cycle. */
        if (!strcasecmp(event,"eviction-del")) {
            advise_large_objects = 1;
            advices++;
        }

        if (!strcasecmp(event,"eviction-cycle")) {
            advise_mass_eviction = 1;
            advices++;
        }

        report = sdscatlen(report,"\n",1);
    }
    dictReleaseIterator(di);

    /* Add non event based advices. */
    if (THPGetAnonHugePagesSize() > 0) {
        advise_disable_thp = 1;
        advices++;
    }

    if (eventnum == 0 && advices == 0) {
        report = sdscat(report,"Dave, no latency spike was observed during the lifetime of this Redis instance, not in the slightest bit. I honestly think you ought to sit down calmly, take a stress pill, and think things over.\n");
    } else if (eventnum > 0 && advices == 0) {
        report = sdscat(report,"\nWhile there are latency events logged, I'm not able to suggest any easy fix. Please use the Redis community to get some help, providing this report in your help request.\n");
    } else {
        /* Add all the suggestions accumulated so far. */

        /* Better VM. */
        report = sdscat(report,"\nI have a few advices for you:\n\n");
        if (advise_better_vm) {
            report = sdscat(report,"- If you are using a virtual machine, consider upgrading it with a faster one using a hypervisior that provides less latency during fork() calls. Xen is known to have poor fork() performance. Even in the context of the same VM provider, certain kinds of instances can execute fork faster than others.\n");
        }

        /* Slow log. */
        if (advise_slowlog_enabled) {
            report = sdscatprintf(report,"- There are latency issues with potentially slow commands you are using. Try to enable the Slow Log Redis feature using the command 'CONFIG SET slowlog-log-slower-than %llu'. If the Slow log is disabled Redis is not able to log slow commands execution for you.\n", (unsigned long long)server.latency_monitor_threshold*1000);
        }

        if (advise_slowlog_tuning) {
            report = sdscatprintf(report,"- Your current Slow Log configuration only logs events that are slower than your configured latency monitor threshold. Please use 'CONFIG SET slowlog-log-slower-than %llu'.\n", (unsigned long long)server.latency_monitor_threshold*1000);
        }

        if (advise_slowlog_inspect) {
            report = sdscat(report,"- Check your Slow Log to understand what are the commands you are running which are too slow to execute. Please check https://redis.io/commands/slowlog for more information.\n");
        }

        /* Intrinsic latency. */
        if (advise_scheduler) {
            report = sdscat(report,"- The system is slow to execute Redis code paths not containing system calls. This usually means the system does not provide Redis CPU time to run for long periods. You should try to:\n"
            "  1) Lower the system load.\n"
            "  2) Use a computer / VM just for Redis if you are running other software in the same system.\n"
            "  3) Check if you have a \"noisy neighbour\" problem.\n"
            "  4) Check with 'redis-cli --intrinsic-latency 100' what is the intrinsic latency in your system.\n"
            "  5) Check if the problem is allocator-related by recompiling Redis with MALLOC=libc, if you are using Jemalloc. However this may create fragmentation problems.\n");
        }

        /* AOF / Disk latency. */
        if (advise_local_disk) {
            report = sdscat(report,"- It is strongly advised to use local disks for persistence, especially if you are using AOF. Remote disks provided by platform-as-a-service providers are known to be slow.\n");
        }

        if (advise_ssd) {
            report = sdscat(report,"- SSD disks are able to reduce fsync latency, and total time needed for snapshotting and AOF log rewriting (resulting in smaller memory usage and smaller final AOF rewrite buffer flushes). With extremely high write load SSD disks can be a good option. However Redis should perform reasonably with high load using normal disks. Use this advice as a last resort.\n");
        }

        if (advise_data_writeback) {
            report = sdscat(report,"- Mounting ext3/4 filesystems with data=writeback can provide a performance boost compared to data=ordered, however this mode of operation provides less guarantees, and sometimes it can happen that after a hard crash the AOF file will have a half-written command at the end and will require to be repaired before Redis restarts.\n");
        }

        if (advise_disk_contention) {
            report = sdscat(report,"- Try to lower the disk contention. This is often caused by other disk intensive processes running in the same computer (including other Redis instances).\n");
        }

        if (advise_no_appendfsync) {
            report = sdscat(report,"- Assuming from the point of view of data safety this is viable in your environment, you could try to enable the 'no-appendfsync-on-rewrite' option, so that fsync will not be performed while there is a child rewriting the AOF file or producing an RDB file (the moment where there is high disk contention).\n");
        }

        if (advise_relax_fsync_policy && server.aof_fsync == AOF_FSYNC_ALWAYS) {
            report = sdscat(report,"- Your fsync policy is set to 'always'. It is very hard to get good performances with such a setup, if possible try to relax the fsync policy to 'onesec'.\n");
        }

        if (advise_write_load_info) {
            report = sdscat(report,"- Latency during the AOF atomic rename operation or when the final difference is flushed to the AOF file at the end of the rewrite, sometimes is caused by very high write load, causing the AOF buffer to get very large. If possible try to send less commands to accomplish the same work, or use Lua scripts to group multiple operations into a single EVALSHA call.\n");
        }

        if (advise_hz && server.hz < 100) {
            report = sdscat(report,"- In order to make the Redis keys expiring process more incremental, try to set the 'hz' configuration parameter to 100 using 'CONFIG SET hz 100'.\n");
        }

        if (advise_large_objects) {
            report = sdscat(report,"- Deleting, expiring or evicting (because of maxmemory policy) large objects is a blocking operation. If you have very large objects that are often deleted, expired, or evicted, try to fragment those objects into multiple smaller objects.\n");
        }

        if (advise_mass_eviction) {
            report = sdscat(report,"- Sudden changes to the 'maxmemory' setting via 'CONFIG SET', or allocation of large objects via sets or sorted sets intersections, STORE option of SORT, Redis Cluster large keys migrations (RESTORE command), may create sudden memory pressure forcing the server to block trying to evict keys. \n");
        }

        if (advise_disable_thp) {
            report = sdscat(report,"- I detected a non zero amount of anonymous huge pages used by your process. This creates very serious latency events in different conditions, especially when Redis is persisting on disk. To disable THP support use the command 'echo never > /sys/kernel/mm/transparent_hugepage/enabled', make sure to also add it into /etc/rc.local so that the command will be executed again after a reboot. Note that even if you have already disabled THP, you still need to restart the Redis process to get rid of the huge pages already created.\n");
        }
    }

    return report;
}

/* ---------------------- Latency command implementation -------------------- */

/* latencyCommand() helper to produce a time-delay reply for all the samples
 * in memory for the specified time series. */
void latencyCommandReplyWithSamples(client *c, struct latencyTimeSeries *ts) {
    void *replylen = addReplyDeferredLen(c);
    int samples = 0, j;

    for (j = 0; j < LATENCY_TS_LEN; j++) {
        int i = (ts->idx + j) % LATENCY_TS_LEN;

        if (ts->samples[i].time == 0) continue;
        addReplyArrayLen(c,2);
        addReplyLongLong(c,ts->samples[i].time);
        addReplyLongLong(c,ts->samples[i].latency);
        samples++;
    }
    setDeferredArrayLen(c,replylen,samples);
}

/* latencyCommand() helper to produce the reply for the LATEST subcommand,
 * listing the last latency sample for every event type registered so far. */
void latencyCommandReplyWithLatestEvents(client *c) {
    dictIterator *di;
    dictEntry *de;

    addReplyArrayLen(c,dictSize(server.latency_events));
    di = dictGetIterator(server.latency_events);
    while((de = dictNext(di)) != NULL) {
        char *event = dictGetKey(de);
        struct latencyTimeSeries *ts = dictGetVal(de);
        int last = (ts->idx + LATENCY_TS_LEN - 1) % LATENCY_TS_LEN;

        addReplyArrayLen(c,4);
        addReplyBulkCString(c,event);
        addReplyLongLong(c,ts->samples[last].time);
        addReplyLongLong(c,ts->samples[last].latency);
        addReplyLongLong(c,ts->max);
    }
    dictReleaseIterator(di);
}

#define LATENCY_GRAPH_COLS 80
sds latencyCommandGenSparkeline(char *event, struct latencyTimeSeries *ts) {
    int j;
    struct sequence *seq = createSparklineSequence();
    sds graph = sdsempty();
    uint32_t min = 0, max = 0;

    for (j = 0; j < LATENCY_TS_LEN; j++) {
        int i = (ts->idx + j) % LATENCY_TS_LEN;
        int elapsed;
        char buf[64];

        if (ts->samples[i].time == 0) continue;
        /* Update min and max. */
        if (seq->length == 0) {
            min = max = ts->samples[i].latency;
        } else {
            if (ts->samples[i].latency > max) max = ts->samples[i].latency;
            if (ts->samples[i].latency < min) min = ts->samples[i].latency;
        }
        /* Use as label the number of seconds / minutes / hours / days
         * ago the event happened. */
        elapsed = time(NULL) - ts->samples[i].time;
        if (elapsed < 60)
            snprintf(buf,sizeof(buf),"%ds",elapsed);
        else if (elapsed < 3600)
            snprintf(buf,sizeof(buf),"%dm",elapsed/60);
        else if (elapsed < 3600*24)
            snprintf(buf,sizeof(buf),"%dh",elapsed/3600);
        else
            snprintf(buf,sizeof(buf),"%dd",elapsed/(3600*24));
        sparklineSequenceAddSample(seq,ts->samples[i].latency,buf);
    }

    graph = sdscatprintf(graph,
        "%s - high %lu ms, low %lu ms (all time high %lu ms)\n", event,
        (unsigned long) max, (unsigned long) min, (unsigned long) ts->max);
    for (j = 0; j < LATENCY_GRAPH_COLS; j++)
        graph = sdscatlen(graph,"-",1);
    graph = sdscatlen(graph,"\n",1);
    graph = sparklineRender(graph,seq,LATENCY_GRAPH_COLS,4,SPARKLINE_FILL);
    freeSparklineSequence(seq);
    return graph;
}

/* LATENCY command implementations.
 *
 * LATENCY HISTORY: return time-latency samples for the specified event.
 * LATENCY LATEST: return the latest latency for all the events classes.
 * LATENCY DOCTOR: returns a human readable analysis of instance latency.
 * LATENCY GRAPH: provide an ASCII graph of the latency of the specified event.
 * LATENCY RESET: reset data of a specified event or all the data if no event provided.
 */
void latencyCommand(client *c) {
    struct latencyTimeSeries *ts;

    if (!strcasecmp(c->argv[1]->ptr,"history") && c->argc == 3) {
        /* LATENCY HISTORY <event> */
        ts = dictFetchValue(server.latency_events,c->argv[2]->ptr);
        if (ts == NULL) {
            addReplyArrayLen(c,0);
        } else {
            latencyCommandReplyWithSamples(c,ts);
        }
    } else if (!strcasecmp(c->argv[1]->ptr,"graph") && c->argc == 3) {
        /* LATENCY GRAPH <event> */
        sds graph;
        dictEntry *de;
        char *event;

        de = dictFind(server.latency_events,c->argv[2]->ptr);
        if (de == NULL) goto nodataerr;
        ts = dictGetVal(de);
        event = dictGetKey(de);

        graph = latencyCommandGenSparkeline(event,ts);
        addReplyVerbatim(c,graph,sdslen(graph),"txt");
        sdsfree(graph);
    } else if (!strcasecmp(c->argv[1]->ptr,"latest") && c->argc == 2) {
        /* LATENCY LATEST */
        latencyCommandReplyWithLatestEvents(c);
    } else if (!strcasecmp(c->argv[1]->ptr,"doctor") && c->argc == 2) {
        /* LATENCY DOCTOR */
        sds report = createLatencyReport();

        addReplyVerbatim(c,report,sdslen(report),"txt");
        sdsfree(report);
    } else if (!strcasecmp(c->argv[1]->ptr,"reset") && c->argc >= 2) {
        /* LATENCY RESET */
        if (c->argc == 2) {
            addReplyLongLong(c,latencyResetEvent(NULL));
        } else {
            int j, resets = 0;

            for (j = 2; j < c->argc; j++)
                resets += latencyResetEvent(c->argv[j]->ptr);
            addReplyLongLong(c,resets);
        }
    } else if (!strcasecmp(c->argv[1]->ptr,"help") && c->argc == 2) {
        const char *help[] = {
"DOCTOR",
"    Return a human readable latency analysis report.",
"GRAPH <event>",
"    Return an ASCII latency graph for the <event> class.",
"HISTORY <event>",
"    Return time-latency samples for the <event> class.",
"LATEST",
"    Return the latest latency samples for all events.",
"RESET [<event> ...]",
"    Reset latency data of one or more <event> classes.",
"    (default: reset all data for all event classes)",
NULL
        };
        addReplyHelp(c, help);
    } else {
        addReplySubcommandSyntaxError(c);
    }
    return;

nodataerr:
    /* Common error when the user asks for an event we have no latency
     * information about. */
    addReplyErrorFormat(c,
        "No samples available for event '%s'", (char*) c->argv[2]->ptr);
}