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     struct timeval timeout = {
 151             .tv_sec = 0,
 152             .tv_usec = 0
 153     };
 154     uint32_t options;
 155     int enabled             = config_get_boolean(CONFIG_SECTION_BACKEND, "enabled", 0);
 156     const char *source      = config_get(CONFIG_SECTION_BACKEND, "data source", "average");
 157     const char *type        = config_get(CONFIG_SECTION_BACKEND, "type", "graphite");
 158     const char *destination = config_get(CONFIG_SECTION_BACKEND, "destination", "localhost");
 159     const char *prefix      = config_get(CONFIG_SECTION_BACKEND, "prefix", "netdata");
 160     const char *hostname    = config_get(CONFIG_SECTION_BACKEND, "hostname", localhost->hostname);
 161     int frequency           = (int)config_get_number(CONFIG_SECTION_BACKEND, "update every", 10);
 162     int buffer_on_failures  = (int)config_get_number(CONFIG_SECTION_BACKEND, "buffer on failures", 10);
 163     long timeoutms          = config_get_number(CONFIG_SECTION_BACKEND, "timeout ms", frequency * 2 * 1000);
 164
 165     // ------------------------------------------------------------------------
 166     // validate configuration options
 167     // and prepare for sending data to our backend
 168
 169     if(!enabled || frequency < 1)
 170         goto cleanup;
 171
 172     if(!strcmp(source, "as collected")) {
 173         options = BACKEND_SOURCE_DATA_AS_COLLECTED;
 174     }
 175     else if(!strcmp(source, "average")) {
 176         options = BACKEND_SOURCE_DATA_AVERAGE;
 177     }
 178     else if(!strcmp(source, "sum") || !strcmp(source, "volume")) {
 179         options = BACKEND_SOURCE_DATA_SUM;
 180     }
 181     else {
 182         error("Invalid data source method '%s' for backend given. Disabling backed.", source);
 183         goto cleanup;
 184     }
 185
 186     if(timeoutms < 1) {
 187         error("BACKED invalid timeout %ld ms given. Assuming %d ms.", timeoutms, frequency * 2 * 1000);
 188         timeoutms = frequency * 2 * 1000;
 189     }
 190     timeout.tv_sec  = (timeoutms * 1000) / 1000000;
 191     timeout.tv_usec = (timeoutms * 1000) % 1000000;
 192
 193
 194     // ------------------------------------------------------------------------
 195     // select the backend type
 196
 197     if(!strcmp(type, "graphite") || !strcmp(type, "graphite:plaintext")) {
 198         default_port = 2003;
 199         if(options == BACKEND_SOURCE_DATA_AS_COLLECTED)
 200             backend_request_formatter = format_dimension_collected_graphite_plaintext;
 201         else
 202             backend_request_formatter = format_dimension_stored_graphite_plaintext;
 203
 204         backend_response_checker = process_graphite_response;
 205     }
 206     else if(!strcmp(type, "opentsdb") || !strcmp(type, "opentsdb:telnet")) {
 207         default_port = 4242;
 208         if(options == BACKEND_SOURCE_DATA_AS_COLLECTED)
 209             backend_request_formatter = format_dimension_collected_opentsdb_telnet;
 210         else
 211             backend_request_formatter = format_dimension_stored_opentsdb_telnet;
 212
 213         backend_response_checker = process_opentsdb_response;
 214     }
 215     else {
 216         error("Unknown backend type '%s'", type);
 217         goto cleanup;
 218     }
 219
 220     if(backend_request_formatter == NULL || backend_response_checker == NULL) {
 221         error("backend is misconfigured - disabling it.");
 222         goto cleanup;
 223     }
 224
 225
 226     // ------------------------------------------------------------------------
 227     // prepare the charts for monitoring the backend operation
 228
 229     struct rusage thread;
 230
 231     collected_number
 232             chart_buffered_metrics = 0,
 233             chart_lost_metrics = 0,
 234             chart_sent_metrics = 0,
 235             chart_buffered_bytes = 0,
 236             chart_received_bytes = 0,
 237             chart_sent_bytes = 0,
 238             chart_receptions = 0,
 239             chart_transmission_successes = 0,
 240             chart_transmission_failures = 0,
 241             chart_data_lost_events = 0,
 242             chart_lost_bytes = 0,
 243             chart_backend_reconnects = 0,
 244             chart_backend_latency = 0;
 245
 246     RRDSET *chart_metrics = rrdset_create_localhost("netdata", "backend_metrics", NULL, "backend", NULL, "Netdata Buffered Metrics", "metrics", 130600, frequency, RRDSET_TYPE_LINE);
 247     rrddim_add(chart_metrics, "buffered", NULL,  1, 1, RRD_ALGORITHM_ABSOLUTE);
 248     rrddim_add(chart_metrics, "lost",     NULL,  1, 1, RRD_ALGORITHM_ABSOLUTE);
 249     rrddim_add(chart_metrics, "sent",     NULL,  1, 1, RRD_ALGORITHM_ABSOLUTE);
 250
 251     RRDSET *chart_bytes = rrdset_create_localhost("netdata", "backend_bytes", NULL, "backend", NULL, "Netdata Backend Data Size", "KB", 130610, frequency, RRDSET_TYPE_AREA);
 252     rrddim_add(chart_bytes, "buffered", NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 253     rrddim_add(chart_bytes, "lost",     NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 254     rrddim_add(chart_bytes, "sent",     NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 255     rrddim_add(chart_bytes, "received", NULL, 1, 1024, RRD_ALGORITHM_ABSOLUTE);
 256
 257     RRDSET *chart_ops = rrdset_create_localhost("netdata", "backend_ops", NULL, "backend", NULL, "Netdata Backend Operations", "operations", 130630, frequency, RRDSET_TYPE_LINE);
 258     rrddim_add(chart_ops, "write",     NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 259     rrddim_add(chart_ops, "discard",   NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 260     rrddim_add(chart_ops, "reconnect", NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 261     rrddim_add(chart_ops, "failure",   NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 262     rrddim_add(chart_ops, "read",      NULL, 1, 1, RRD_ALGORITHM_ABSOLUTE);
 263
 264     /*
 265      * this is misleading - we can only measure the time we need to send data
 266      * this time is not related to the time required for the data to travel to
 267      * the backend database and the time that server needed to process them
 268      *
 269      * issue #1432 and https://www.softlab.ntua.gr/facilities/documentation/unix/unix-socket-faq/unix-socket-faq-2.html
 270      *
 271     RRDSET *chart_latency = rrdset_create_localhost("netdata", "backend_latency", NULL, "backend", NULL, "Netdata Backend Latency", "ms", 130620, frequency, RRDSET_TYPE_AREA);
 272     rrddim_add(chart_latency, "latency",   NULL,  1, 1000, RRD_ALGORITHM_ABSOLUTE);
 273     */
 274
 275     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);
 276     rrddim_add(chart_rusage, "user",   NULL, 1, 1000, RRD_ALGORITHM_INCREMENTAL);
 277     rrddim_add(chart_rusage, "system", NULL, 1, 1000, RRD_ALGORITHM_INCREMENTAL);
 278
 279
 280     // ------------------------------------------------------------------------
 281     // prepare the backend main loop
 282
 283     info("BACKEND configured ('%s' on '%s' sending '%s' data, every %d seconds, as host '%s', with prefix '%s')", type, destination, source, frequency, hostname, prefix);
 284
 285     usec_t step_ut = frequency * USEC_PER_SEC;
 286     time_t after = now_realtime_sec();
 287     int failures = 0;
 288     heartbeat_t hb;
 289     heartbeat_init(&hb);
 290
 291     for(;;) {
 292
 293         // ------------------------------------------------------------------------
 294         // Wait for the next iteration point.
 295         heartbeat_next(&hb, step_ut);
 296         time_t before = now_realtime_sec();
 297
 298
 299         // ------------------------------------------------------------------------
 300         // add to the buffer the data we need to send to the backend
 301
 302         int pthreadoldcancelstate;
 303
 304         if(unlikely(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &pthreadoldcancelstate) != 0))
 305             error("Cannot set pthread cancel state to DISABLE.");
 306
 307         rrd_rdlock();
 308         RRDHOST *host;
 309         rrdhost_foreach_read(host) {
 310             if(host->rrd_memory_mode == RRD_MEMORY_MODE_NONE)
 311                 continue;
 312
 313             rrdhost_rdlock(host);
 314
 315             RRDSET *st;
 316             rrdset_foreach_read(st, host) {
 317                 rrdset_rdlock(st);
 318
 319                 RRDDIM *rd;
 320                 rrddim_foreach_read(rd, st) {
 321                     if(rd->last_collected_time.tv_sec >= after)
 322                         chart_buffered_metrics += backend_request_formatter(b, prefix, host, (host == localhost)?hostname:host->hostname, st, rd, after, before, options);
 323                 }
 324                 rrdset_unlock(st);
 325             }
 326             rrdhost_unlock(host);
 327         }
 328         rrd_unlock();
 329
 330         if(unlikely(pthread_setcancelstate(pthreadoldcancelstate, NULL) != 0))
 331             error("Cannot set pthread cancel state to RESTORE (%d).", pthreadoldcancelstate);
 332
 333         // ------------------------------------------------------------------------
 334
 335         chart_buffered_bytes = (collected_number)buffer_strlen(b);
 336
 337         // reset the monitoring chart counters
 338         chart_received_bytes =
 339         chart_sent_bytes =
 340         chart_sent_metrics =
 341         chart_lost_metrics =
 342         chart_transmission_successes =
 343         chart_transmission_failures =
 344         chart_data_lost_events =
 345         chart_lost_bytes =
 346         chart_backend_reconnects =
 347         chart_backend_latency = 0;
 348
 349         if(unlikely(netdata_exit)) break;
 350
 351         //fprintf(stderr, "\nBACKEND BEGIN:\n%s\nBACKEND END\n", buffer_tostring(b)); // FIXME
 352         //fprintf(stderr, "after = %lu, before = %lu\n", after, before);
 353
 354         // prepare for the next iteration
 355         // to add incrementally data to buffer
 356         after = before;
 357
 358         // ------------------------------------------------------------------------
 359         // if we are connected, receive a response, without blocking
 360
 361         if(likely(sock != -1)) {
 362             errno = 0;
 363
 364             // loop through to collect all data
 365             while(sock != -1 && errno != EWOULDBLOCK) {
 366                 buffer_need_bytes(response, 4096);
 367
 368                 ssize_t r = recv(sock, &response->buffer[response->len], response->size - response->len, MSG_DONTWAIT);
 369                 if(likely(r > 0)) {
 370                     // we received some data
 371                     response->len += r;
 372                     chart_received_bytes += r;
 373                     chart_receptions++;
 374                 }
 375                 else if(r == 0) {
 376                     error("Backend '%s' closed the socket", destination);
 377                     close(sock);
 378                     sock = -1;
 379                 }
 380                 else {
 381                     // failed to receive data
 382                     if(errno != EAGAIN && errno != EWOULDBLOCK) {
 383                         error("Cannot receive data from backend '%s'.", destination);
 384                     }
 385                 }
 386             }
 387
 388             // if we received data, process them
 389             if(buffer_strlen(response))
 390                 backend_response_checker(response);
 391         }
 392
 393         // ------------------------------------------------------------------------
 394         // if we are not connected, connect to a backend server
 395
 396         if(unlikely(sock == -1)) {
 397             usec_t start_ut = now_monotonic_usec();
 398             size_t reconnects = 0;
 399
 400             sock = connect_to_one_of(destination, default_port, &timeout, &reconnects, NULL, 0);
 401
 402             chart_backend_reconnects += reconnects;
 403             chart_backend_latency += now_monotonic_usec() - start_ut;
 404         }
 405
 406         if(unlikely(netdata_exit)) break;
 407
 408         // ------------------------------------------------------------------------
 409         // if we are connected, send our buffer to the backend server
 410
 411         if(likely(sock != -1)) {
 412             size_t len = buffer_strlen(b);
 413             usec_t start_ut = now_monotonic_usec();
 414             int flags = 0;
 415 #ifdef MSG_NOSIGNAL
 416             flags += MSG_NOSIGNAL;
 417 #endif
 418
 419             ssize_t written = send(sock, buffer_tostring(b), len, flags);
 420             chart_backend_latency += now_monotonic_usec() - start_ut;
 421             if(written != -1 && (size_t)written == len) {
 422                 // we sent the data successfully
 423                 chart_transmission_successes++;
 424                 chart_sent_bytes += written;
 425                 chart_sent_metrics = chart_buffered_metrics;
 426
 427                 // reset the failures count
 428                 failures = 0;
 429
 430                 // empty the buffer
 431                 buffer_flush(b);
 432             }
 433             else {
 434                 // oops! we couldn't send (all or some of the) data
 435                 error("Failed to write data to database backend '%s'. Willing to write %zu bytes, wrote %zd bytes. Will re-connect.", destination, len, written);
 436                 chart_transmission_failures++;
 437
 438                 if(written != -1)
 439                     chart_sent_bytes += written;
 440
 441                 // increment the counter we check for data loss
 442                 failures++;
 443
 444                 // close the socket - we will re-open it next time
 445                 close(sock);
 446                 sock = -1;
 447             }
 448         }
 449         else {
 450             error("Failed to update database backend '%s'", destination);
 451             chart_transmission_failures++;
 452
 453             // increment the counter we check for data loss
 454             failures++;
 455         }
 456
 457         if(failures > buffer_on_failures) {
 458             // too bad! we are going to lose data
 459             chart_lost_bytes += buffer_strlen(b);
 460             error("Reached %d backend failures. Flushing buffers to protect this host - this results in data loss on back-end server '%s'", failures, destination);
 461             buffer_flush(b);
 462             failures = 0;
 463             chart_data_lost_events++;
 464             chart_lost_metrics = chart_buffered_metrics;
 465         }
 466
 467         if(unlikely(netdata_exit)) break;
 468
 469         // ------------------------------------------------------------------------
 470         // update the monitoring charts
 471
 472         if(likely(chart_ops->counter_done)) rrdset_next(chart_ops);
 473         rrddim_set(chart_ops, "read",         chart_receptions);
 474         rrddim_set(chart_ops, "write",        chart_transmission_successes);
 475         rrddim_set(chart_ops, "discard",      chart_data_lost_events);
 476         rrddim_set(chart_ops, "failure",      chart_transmission_failures);
 477         rrddim_set(chart_ops, "reconnect",    chart_backend_reconnects);
 478         rrdset_done(chart_ops);
 479
 480         if(likely(chart_metrics->counter_done)) rrdset_next(chart_metrics);
 481         rrddim_set(chart_metrics, "buffered", chart_buffered_metrics);
 482         rrddim_set(chart_metrics, "lost",     chart_lost_metrics);
 483         rrddim_set(chart_metrics, "sent",     chart_sent_metrics);
 484         rrdset_done(chart_metrics);
 485
 486         if(likely(chart_bytes->counter_done)) rrdset_next(chart_bytes);
 487         rrddim_set(chart_bytes, "buffered",   chart_buffered_bytes);
 488         rrddim_set(chart_bytes, "lost",       chart_lost_bytes);
 489         rrddim_set(chart_bytes, "sent",       chart_sent_bytes);
 490         rrddim_set(chart_bytes, "received",   chart_received_bytes);
 491         rrdset_done(chart_bytes);
 492
 493         /*
 494         if(likely(chart_latency->counter_done)) rrdset_next(chart_latency);
 495         rrddim_set(chart_latency, "latency",  chart_backend_latency);
 496         rrdset_done(chart_latency);
 497         */
 498
 499         getrusage(RUSAGE_THREAD, &thread);
 500         if(likely(chart_rusage->counter_done)) rrdset_next(chart_rusage);
 501         rrddim_set(chart_rusage, "user",   thread.ru_utime.tv_sec * 1000000ULL + thread.ru_utime.tv_usec);
 502         rrddim_set(chart_rusage, "system", thread.ru_stime.tv_sec * 1000000ULL + thread.ru_stime.tv_usec);
 503         rrdset_done(chart_rusage);
 504
 505         if(likely(buffer_strlen(b) == 0))
 506             chart_buffered_metrics = 0;
 507
 508         if(unlikely(netdata_exit)) break;
 509     }
 510
 511 cleanup:
 512     if(sock != -1)
 513         close(sock);
 514
 515     buffer_free(b);
 516     buffer_free(response);
 517
 518     info("BACKEND thread exiting");
 519
 520     static_thread->enabled = 0;
 521     pthread_exit(NULL);
 522     return NULL;
 523 }