src/backends.c

   1 #include "common.h"
   2
   3 #define BACKEND_SOURCE_DATA_AS_COLLECTED 0x00000001
   4 #define BACKEND_SOURCE_DATA_AVERAGE      0x00000002
   5 #define BACKEND_SOURCE_DATA_SUM          0x00000004
   6
   7 static inline calculated_number backend_calculate_value_from_stored_data(RRDSET *st, RRDDIM *rd, time_t after, time_t before, uint32_t options) {
   8     time_t first_t = rrdset_first_entry_t(st);
   9     time_t last_t = rrdset_last_entry_t(st);
  10
  11     if(unlikely(before < first_t || after > last_t))
  12         // the chart has not been updated in the wanted timeframe
  13         return NAN;
  14
  15     // align the time-frame
  16     // for 'after' also skip the first value by adding st->update_every
  17     after  = after  - after  % st->update_every + st->update_every;
  18     before = before - before % st->update_every;
  19
  20     if(unlikely(after < first_t))
  21         after = first_t;
  22
  23     if(unlikely(after > before))
  24         // this can happen when the st->update_every > before - after
  25         before = after;
  26
  27     if(unlikely(before > last_t))
  28         before = last_t;
  29
  30     size_t counter = 0;
  31     calculated_number sum = 0;
  32
  33     long    start_at_slot = rrdset_time2slot(st, before),
  34             stop_at_slot  = rrdset_time2slot(st, after),
  35             slot, stop_now = 0;
  36
  37     for(slot = start_at_slot; !stop_now ; slot--) {
  38         if(unlikely(slot < 0)) slot = st->entries - 1;
  39         if(unlikely(slot == stop_at_slot)) stop_now = 1;
  40
  41         storage_number n = rd->values[slot];
  42         if(unlikely(!does_storage_number_exist(n))) continue;
  43
  44         calculated_number value = unpack_storage_number(n);
  45         sum += value;
  46         counter++;
  47     }
  48
  49     if(unlikely(!counter))
  50         return NAN;
  51
  52     if(unlikely(options & BACKEND_SOURCE_DATA_SUM))
  53         return sum;
  54
  55     return sum / (calculated_number)counter;
  56 }
  57
  58 static inline int format_dimension_collected_graphite_plaintext(BUFFER *b, const char *prefix, RRDHOST *host, const char *hostname, RRDSET *st, RRDDIM *rd, time_t after, time_t before, uint32_t options) {
  59     (void)host;
  60     (void)after;
  61     (void)before;
  62     (void)options;
  63     buffer_sprintf(b, "%s.%s.%s.%s " COLLECTED_NUMBER_FORMAT " %u\n", prefix, hostname, st->id, rd->id, rd->last_collected_value, (uint32_t)rd->last_collected_time.tv_sec);
  64     return 1;
  65 }
  66
  67 static inline int format_dimension_stored_graphite_plaintext(BUFFER *b, const char *prefix, RRDHOST *host, const char *hostname, RRDSET *st, RRDDIM *rd, time_t after, time_t before, uint32_t options) {
  68     (void)host;
  69     calculated_number value = backend_calculate_value_from_stored_data(st, rd, after, before, options);
  70     if(!isnan(value)) {
  71         buffer_sprintf(b, "%s.%s.%s.%s " CALCULATED_NUMBER_FORMAT " %u\n", prefix, hostname, st->id, rd->id, value, (uint32_t) before);
  72         return 1;
  73     }
  74     return 0;
  75 }
  76
  77 static inline int format_dimension_collected_opentsdb_telnet(BUFFER *b, const char *prefix, RRDHOST *host, const char *hostname, RRDSET *st, RRDDIM *rd, time_t after, time_t before, uint32_t options) {
  78     (void)host;
  79     (void)after;
  80     (void)before;
  81     (void)options;
  82     buffer_sprintf(b, "put %s.%s.%s %u " COLLECTED_NUMBER_FORMAT " host=%s\n", prefix, st->id, rd->id, (uint32_t)rd->last_collected_time.tv_sec, rd->last_collected_value, hostname);
  83     return 1;
  84 }
  85
  86 static inline int format_dimension_stored_opentsdb_telnet(BUFFER *b, const char *prefix, RRDHOST *host, const char *hostname, RRDSET *st, RRDDIM *rd, time_t after, time_t before, uint32_t options) {
  87     (void)host;
  88     calculated_number value = backend_calculate_value_from_stored_data(st, rd, after, before, options);
  89     if(!isnan(value)) {
  90         buffer_sprintf(b, "put %s.%s.%s %u " CALCULATED_NUMBER_FORMAT " host=%s\n", prefix, st->id, rd->id, (uint32_t) before, value, hostname);
  91         return 1;
  92     }
  93     return 0;
  94 }
  95
  96 static inline int process_graphite_response(BUFFER *b) {
  97     char sample[1024];
  98     const char *s = buffer_tostring(b);
  99     char *d = sample, *e = &sample[sizeof(sample) - 1];
 100
 101     for(; *s && d < e ;s++) {
 102         char c = *s;
 103         if(unlikely(!isprint(c))) c = ' ';
 104         *d++ = c;
 105     }
 106     *d = '\0';
 107
 108     info("Received %zu bytes from graphite backend. Ignoring them. Sample: '%s'", buffer_strlen(b), sample);
 109     buffer_flush(b);
 110     return 0;
 111 }
 112
 113 static inline int process_opentsdb_response(BUFFER *b) {
 114     char sample[1024];
 115     const char *s = buffer_tostring(b);
 116     char *d = sample, *e = &sample[sizeof(sample) - 1];
 117
 118     for(; *s && d < e ;s++) {
 119         char c = *s;
 120         if(unlikely(!isprint(c))) c = ' ';
 121         *d++ = c;
 122     }
 123     *d = '\0';
 124
 125     info("Received %zu bytes from opentsdb backend. Ignoring them. Sample: '%s'", buffer_strlen(b), sample);
 126     buffer_flush(b);
 127     return 0;
 128 }
 129
 130 void *backends_main(void *ptr) {
 131     int default_port = 0;
 132     int sock = -1;
 133     struct netdata_static_thread *static_thread = (struct netdata_static_thread *)ptr;
 134
 135     BUFFER *b = buffer_create(1), *response = buffer_create(1);
 136     int (*backend_request_formatter)(BUFFER *, const char *, RRDHOST *, const char *, RRDSET *, RRDDIM *, time_t, time_t, uint32_t) = NULL;
 137     int (*backend_response_checker)(BUFFER *) = NULL;
 138
 139     info("BACKEND thread created with task id %d", gettid());
 140
 141     if(pthread_setcanceltype(PTHREAD_CANCEL_DEFERRED, NULL) != 0)
 142         error("Cannot set pthread cancel type to DEFERRED.");
 143
 144     if(pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL) != 0)
 145         error("Cannot set pthread cancel state to ENABLE.");
 146
 147     // ------------------------------------------------------------------------
 148     // collect configuration options
 149
 150     if(rrdpush_exclusive) {
 151         info("Backend is disabled - use the central netdata");
 152         goto cleanup;
 153     }
 154
 155     struct timeval timeout = {
 156             .tv_sec = 0,
 157             .tv_usec = 0
 158     };
 159     uint32_t options;
 160     int enabled             = config_get_boolean("backend", "enabled", 0);
 161     const char *source      = config_get("backend", "data source", "average");
 162     const char *type        = config_get("backend", "type", "graphite");
 163     const char *destination = config_get("backend", "destination", "localhost");
 164     const char *prefix      = config_get("backend", "prefix", "netdata");
 165     const char *hostname    = config_get("backend", "hostname", localhost->hostname);
 166     int frequency           = (int)config_get_number("backend", "update every", 10);
 167     int buffer_on_failures  = (int)config_get_number("backend", "buffer on failures", 10);
 168     long timeoutms          = config_get_number("backend", "timeout ms", frequency * 2 * 1000);
 169
 170     // ------------------------------------------------------------------------
 171     // validate configuration options
 172     // and prepare for sending data to our backend
 173
 174     if(!enabled || frequency < 1)
 175         goto cleanup;
 176
 177     if(!strcmp(source, "as collected")) {
 178         options = BACKEND_SOURCE_DATA_AS_COLLECTED;
 179     }
 180     else if(!strcmp(source, "average")) {
 181         options = BACKEND_SOURCE_DATA_AVERAGE;
 182     }
 183     else if(!strcmp(source, "sum") || !strcmp(source, "volume")) {
 184         options = BACKEND_SOURCE_DATA_SUM;
 185     }
 186     else {
 187         error("Invalid data source method '%s' for backend given. Disabling backed.", source);
 188         goto cleanup;
 189     }
 190
 191     if(timeoutms < 1) {
 192         error("BACKED invalid timeout %ld ms given. Assuming %d ms.", timeoutms, frequency * 2 * 1000);
 193         timeoutms = frequency * 2 * 1000;
 194     }
 195     timeout.tv_sec  = (timeoutms * 1000) / 1000000;
 196     timeout.tv_usec = (timeoutms * 1000) % 1000000;
 197
 198
 199     // ------------------------------------------------------------------------
 200     // select the backend type
 201
 202     if(!strcmp(type, "graphite") || !strcmp(type, "graphite:plaintext")) {
 203         default_port = 2003;
 204         if(options == BACKEND_SOURCE_DATA_AS_COLLECTED)
 205             backend_request_formatter = format_dimension_collected_graphite_plaintext;
 206         else
 207             backend_request_formatter = format_dimension_stored_graphite_plaintext;
 208
 209         backend_response_checker = process_graphite_response;
 210     }
 211     else if(!strcmp(type, "opentsdb") || !strcmp(type, "opentsdb:telnet")) {
 212         default_port = 4242;
 213         if(options == BACKEND_SOURCE_DATA_AS_COLLECTED)
 214             backend_request_formatter = format_dimension_collected_opentsdb_telnet;
 215         else
 216             backend_request_formatter = format_dimension_stored_opentsdb_telnet;
 217
 218         backend_response_checker = process_opentsdb_response;
 219     }
 220     else {
 221         error("Unknown backend type '%s'", type);
 222         goto cleanup;
 223     }
 224
 225     if(backend_request_formatter == NULL || backend_response_checker == NULL) {
 226         error("backend is misconfigured - disabling it.");
 227         goto cleanup;
 228     }
 229
 230
 231     // ------------------------------------------------------------------------
 232     // prepare the charts for monitoring the backend operation
 233
 234     struct rusage thread;
 235
 236     collected_number
 237             chart_buffered_metrics = 0,
 238             chart_lost_metrics = 0,
 239             chart_sent_metrics = 0,
 240             chart_buffered_bytes = 0,
 241             chart_received_bytes = 0,
 242             chart_sent_bytes = 0,
 243             chart_receptions = 0,
 244             chart_transmission_successes = 0,
 245             chart_transmission_failures = 0,
 246             chart_data_lost_events = 0,
 247             chart_lost_bytes = 0,
 248             chart_backend_reconnects = 0,
 249             chart_backend_latency = 0;
 250
 251     RRDSET *chart_metrics = rrdset_create_localhost("netdata", "backend_metrics", NULL, "backend", NULL, "Netdata Buffered Metrics", "metrics", 130600, frequency, RRDSET_TYPE_LINE);
 252     rrddim_add(chart_metrics, "buffered", NULL,  1, 1, RRD_ALGORITHM_ABSOLUTE);
 253     rrddim_add(chart_metrics, "lost",     NULL,  1, 1, RRD_ALGORITHM_ABSOLUTE);
 254     rrddim_add(chart_metrics, "sent",     NULL,  1, 1, RRD_ALGORITHM_ABSOLUTE);
 255
 256     RRDSET *chart_bytes = rrdset_create_localhost("netdata", "backend_bytes", NULL, "backend", NULL, "Netdata Backend Data Size", "KB", 130610, frequency, RRDSET_TYPE_AREA);
 257     rrddim_add(chart_bytes, "buffered", NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 258     rrddim_add(chart_bytes, "lost",     NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 259     rrddim_add(chart_bytes, "sent",     NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 260     rrddim_add(chart_bytes, "received", NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 261
 262     RRDSET *chart_ops = rrdset_create_localhost("netdata", "backend_ops", NULL, "backend", NULL, "Netdata Backend Operations", "operations", 130630, frequency, RRDSET_TYPE_LINE);
 263     rrddim_add(chart_ops, "write",     NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 264     rrddim_add(chart_ops, "discard",   NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 265     rrddim_add(chart_ops, "reconnect", NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 266     rrddim_add(chart_ops, "failure",   NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 267     rrddim_add(chart_ops, "read",      NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 268
 269     /*
 270      * this is misleading - we can only measure the time we need to send data
 271      * this time is not related to the time required for the data to travel to
 272      * the backend database and the time that server needed to process them
 273      *
 274      * issue #1432 and https://www.softlab.ntua.gr/facilities/documentation/unix/unix-socket-faq/unix-socket-faq-2.html
 275      *
 276     RRDSET *chart_latency = rrdset_create_localhost("netdata", "backend_latency", NULL, "backend", NULL, "Netdata Backend Latency", "ms", 130620, frequency, RRDSET_TYPE_AREA);
 277     rrddim_add(chart_latency, "latency",   NULL,  1, 1000, RRD_ALGORITHM_ABSOLUTE);
 278     */
 279
 280     RRDSET *chart_rusage = rrdset_create_localhost("netdata", "backend_thread_cpu", NULL, "backend", NULL, "NetData Backend Thread CPU usage", "milliseconds/s", 130630, frequency, RRDSET_TYPE_STACKED);
 281     rrddim_add(chart_rusage, "user",   NULL, 1, 1000, RRD_ALGORITHM_INCREMENTAL);
 282     rrddim_add(chart_rusage, "system", NULL, 1, 1000, RRD_ALGORITHM_INCREMENTAL);
 283
 284
 285     // ------------------------------------------------------------------------
 286     // prepare the backend main loop
 287
 288     info("BACKEND configured ('%s' on '%s' sending '%s' data, every %d seconds, as host '%s', with prefix '%s')", type, destination, source, frequency, hostname, prefix);
 289
 290     usec_t step_ut = frequency * USEC_PER_SEC;
 291     time_t after = now_realtime_sec();
 292     int failures = 0;
 293     heartbeat_t hb;
 294     heartbeat_init(&hb);
 295
 296     for(;;) {
 297
 298         // ------------------------------------------------------------------------
 299         // Wait for the next iteration point.
 300         heartbeat_next(&hb, step_ut);
 301         time_t before = now_realtime_sec();
 302
 303
 304         // ------------------------------------------------------------------------
 305         // add to the buffer the data we need to send to the backend
 306
 307         int pthreadoldcancelstate;
 308
 309         if(unlikely(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &pthreadoldcancelstate) != 0))
 310             error("Cannot set pthread cancel state to DISABLE.");
 311
 312         rrd_rdlock();
 313         RRDHOST *host;
 314         rrdhost_foreach_read(host) {
 315             rrdhost_rdlock(host);
 316
 317             RRDSET *st;
 318             rrdset_foreach_read(st, host) {
 319                 rrdset_rdlock(st);
 320
 321                 RRDDIM *rd;
 322                 rrddim_foreach_read(rd, st) {
 323                     if(rd->last_collected_time.tv_sec >= after)
 324                         chart_buffered_metrics += backend_request_formatter(b, prefix, host, (host == localhost)?hostname:host->hostname, st, rd, after, before, options);
 325                 }
 326                 rrdset_unlock(st);
 327             }
 328             rrdhost_unlock(host);
 329         }
 330         rrd_unlock();
 331
 332         if(unlikely(pthread_setcancelstate(pthreadoldcancelstate, NULL) != 0))
 333             error("Cannot set pthread cancel state to RESTORE (%d).", pthreadoldcancelstate);
 334
 335         // ------------------------------------------------------------------------
 336
 337         chart_buffered_bytes = (collected_number)buffer_strlen(b);
 338
 339         // reset the monitoring chart counters
 340         chart_received_bytes =
 341         chart_sent_bytes =
 342         chart_sent_metrics =
 343         chart_lost_metrics =
 344         chart_transmission_successes =
 345         chart_transmission_failures =
 346         chart_data_lost_events =
 347         chart_lost_bytes =
 348         chart_backend_reconnects =
 349         chart_backend_latency = 0;
 350
 351         if(unlikely(netdata_exit)) break;
 352
 353         //fprintf(stderr, "\nBACKEND BEGIN:\n%s\nBACKEND END\n", buffer_tostring(b)); // FIXME
 354         //fprintf(stderr, "after = %lu, before = %lu\n", after, before);
 355
 356         // prepare for the next iteration
 357         // to add incrementally data to buffer
 358         after = before;
 359
 360         // ------------------------------------------------------------------------
 361         // if we are connected, receive a response, without blocking
 362
 363         if(likely(sock != -1)) {
 364             errno = 0;
 365
 366             // loop through to collect all data
 367             while(sock != -1 && errno != EWOULDBLOCK) {
 368                 buffer_need_bytes(response, 4096);
 369
 370                 ssize_t r = recv(sock, &response->buffer[response->len], response->size - response->len, MSG_DONTWAIT);
 371                 if(likely(r > 0)) {
 372                     // we received some data
 373                     response->len += r;
 374                     chart_received_bytes += r;
 375                     chart_receptions++;
 376                 }
 377                 else if(r == 0) {
 378                     error("Backend '%s' closed the socket", destination);
 379                     close(sock);
 380                     sock = -1;
 381                 }
 382                 else {
 383                     // failed to receive data
 384                     if(errno != EAGAIN && errno != EWOULDBLOCK) {
 385                         error("Cannot receive data from backend '%s'.", destination);
 386                     }
 387                 }
 388             }
 389
 390             // if we received data, process them
 391             if(buffer_strlen(response))
 392                 backend_response_checker(response);
 393         }
 394
 395         // ------------------------------------------------------------------------
 396         // if we are not connected, connect to a backend server
 397
 398         if(unlikely(sock == -1)) {
 399             usec_t start_ut = now_monotonic_usec();
 400             size_t reconnects = 0;
 401
 402             sock = connect_to_one_of(destination, default_port, &timeout, &reconnects, NULL, 0);
 403
 404             chart_backend_reconnects += reconnects;
 405             chart_backend_latency += now_monotonic_usec() - start_ut;
 406         }
 407
 408         if(unlikely(netdata_exit)) break;
 409
 410         // ------------------------------------------------------------------------
 411         // if we are connected, send our buffer to the backend server
 412
 413         if(likely(sock != -1)) {
 414             size_t len = buffer_strlen(b);
 415             usec_t start_ut = now_monotonic_usec();
 416             int flags = 0;
 417 #ifdef MSG_NOSIGNAL
 418             flags += MSG_NOSIGNAL;
 419 #endif
 420
 421             ssize_t written = send(sock, buffer_tostring(b), len, flags);
 422             chart_backend_latency += now_monotonic_usec() - start_ut;
 423             if(written != -1 && (size_t)written == len) {
 424                 // we sent the data successfully
 425                 chart_transmission_successes++;
 426                 chart_sent_bytes += written;
 427                 chart_sent_metrics = chart_buffered_metrics;
 428
 429                 // reset the failures count
 430                 failures = 0;
 431
 432                 // empty the buffer
 433                 buffer_flush(b);
 434             }
 435             else {
 436                 // oops! we couldn't send (all or some of the) data
 437                 error("Failed to write data to database backend '%s'. Willing to write %zu bytes, wrote %zd bytes. Will re-connect.", destination, len, written);
 438                 chart_transmission_failures++;
 439
 440                 if(written != -1)
 441                     chart_sent_bytes += written;
 442
 443                 // increment the counter we check for data loss
 444                 failures++;
 445
 446                 // close the socket - we will re-open it next time
 447                 close(sock);
 448                 sock = -1;
 449             }
 450         }
 451         else {
 452             error("Failed to update database backend '%s'", destination);
 453             chart_transmission_failures++;
 454
 455             // increment the counter we check for data loss
 456             failures++;
 457         }
 458
 459         if(failures > buffer_on_failures) {
 460             // too bad! we are going to lose data
 461             chart_lost_bytes += buffer_strlen(b);
 462             error("Reached %d backend failures. Flushing buffers to protect this host - this results in data loss on back-end server '%s'", failures, destination);
 463             buffer_flush(b);
 464             failures = 0;
 465             chart_data_lost_events++;
 466             chart_lost_metrics = chart_buffered_metrics;
 467         }
 468
 469         if(unlikely(netdata_exit)) break;
 470
 471         // ------------------------------------------------------------------------
 472         // update the monitoring charts
 473
 474         if(likely(chart_ops->counter_done)) rrdset_next(chart_ops);
 475         rrddim_set(chart_ops, "read",         chart_receptions);
 476         rrddim_set(chart_ops, "write",        chart_transmission_successes);
 477         rrddim_set(chart_ops, "discard",      chart_data_lost_events);
 478         rrddim_set(chart_ops, "failure",      chart_transmission_failures);
 479         rrddim_set(chart_ops, "reconnect",    chart_backend_reconnects);
 480         rrdset_done(chart_ops);
 481
 482         if(likely(chart_metrics->counter_done)) rrdset_next(chart_metrics);
 483         rrddim_set(chart_metrics, "buffered", chart_buffered_metrics);
 484         rrddim_set(chart_metrics, "lost",     chart_lost_metrics);
 485         rrddim_set(chart_metrics, "sent",     chart_sent_metrics);
 486         rrdset_done(chart_metrics);
 487
 488         if(likely(chart_bytes->counter_done)) rrdset_next(chart_bytes);
 489         rrddim_set(chart_bytes, "buffered",   chart_buffered_bytes);
 490         rrddim_set(chart_bytes, "lost",       chart_lost_bytes);
 491         rrddim_set(chart_bytes, "sent",       chart_sent_bytes);
 492         rrddim_set(chart_bytes, "received",   chart_received_bytes);
 493         rrdset_done(chart_bytes);
 494
 495         /*
 496         if(likely(chart_latency->counter_done)) rrdset_next(chart_latency);
 497         rrddim_set(chart_latency, "latency",  chart_backend_latency);
 498         rrdset_done(chart_latency);
 499         */
 500
 501         getrusage(RUSAGE_THREAD, &thread);
 502         if(likely(chart_rusage->counter_done)) rrdset_next(chart_rusage);
 503         rrddim_set(chart_rusage, "user",   thread.ru_utime.tv_sec * 1000000ULL + thread.ru_utime.tv_usec);
 504         rrddim_set(chart_rusage, "system", thread.ru_stime.tv_sec * 1000000ULL + thread.ru_stime.tv_usec);
 505         rrdset_done(chart_rusage);
 506
 507         if(likely(buffer_strlen(b) == 0))
 508             chart_buffered_metrics = 0;
 509
 510         if(unlikely(netdata_exit)) break;
 511     }
 512
 513 cleanup:
 514     if(sock != -1)
 515         close(sock);
 516
 517     buffer_free(b);
 518     buffer_free(response);
 519
 520     info("BACKEND thread exiting");
 521
 522     static_thread->enabled = 0;
 523     pthread_exit(NULL);
 524     return NULL;
 525 }