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 - after < st->update_every && after != after - after % st->update_every))
  12         // when st->update_every is bigger than the frequency we send data to backend
  13         // skip the iterations that are not aligned to the database
  14         return NAN;
  15
  16     // align the time-frame
  17     // for 'after' also skip the first value by adding st->update_every
  18     after  = after  - after  % st->update_every + st->update_every;
  19     before = before - before % st->update_every;
  20
  21     if(unlikely(after < first_t))
  22         after = first_t;
  23
  24     if(unlikely(after > before))
  25         // this can happen when the st->update_every > before - after
  26         before = after;
  27
  28     if(unlikely(before > last_t))
  29         before = last_t;
  30
  31     size_t counter = 0;
  32     calculated_number sum = 0;
  33
  34     long    start_at_slot = rrdset_time2slot(st, before),
  35             stop_at_slot  = rrdset_time2slot(st, after),
  36             slot, stop_now = 0;
  37
  38     for(slot = start_at_slot; !stop_now ; slot--) {
  39         if(unlikely(slot < 0)) slot = st->entries - 1;
  40         if(unlikely(slot == stop_at_slot)) stop_now = 1;
  41
  42         storage_number n = rd->values[slot];
  43         if(unlikely(!does_storage_number_exist(n))) continue;
  44
  45         calculated_number value = unpack_storage_number(n);
  46         sum += value;
  47         counter++;
  48     }
  49
  50     if(unlikely(!counter))
  51         return NAN;
  52
  53     if(unlikely(options & BACKEND_SOURCE_DATA_SUM))
  54         return sum;
  55
  56     return sum / (calculated_number)counter;
  57 }
  58
  59 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) {
  60     (void)host;
  61     (void)after;
  62     (void)before;
  63     (void)options;
  64     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);
  65     return 1;
  66 }
  67
  68 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) {
  69     (void)host;
  70     calculated_number value = backend_calculate_value_from_stored_data(st, rd, after, before, options);
  71     if(!isnan(value)) {
  72         buffer_sprintf(b, "%s.%s.%s.%s " CALCULATED_NUMBER_FORMAT " %u\n", prefix, hostname, st->id, rd->id, value, (uint32_t) before);
  73         return 1;
  74     }
  75     return 0;
  76 }
  77
  78 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) {
  79     (void)host;
  80     (void)after;
  81     (void)before;
  82     (void)options;
  83     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);
  84     return 1;
  85 }
  86
  87 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) {
  88     (void)host;
  89     calculated_number value = backend_calculate_value_from_stored_data(st, rd, after, before, options);
  90     if(!isnan(value)) {
  91         buffer_sprintf(b, "put %s.%s.%s %u " CALCULATED_NUMBER_FORMAT " host=%s\n", prefix, st->id, rd->id, (uint32_t) before, value, hostname);
  92         return 1;
  93     }
  94     return 0;
  95 }
  96
  97 static inline int process_graphite_response(BUFFER *b) {
  98     char sample[1024];
  99     const char *s = buffer_tostring(b);
 100     char *d = sample, *e = &sample[sizeof(sample) - 1];
 101
 102     for(; *s && d < e ;s++) {
 103         char c = *s;
 104         if(unlikely(!isprint(c))) c = ' ';
 105         *d++ = c;
 106     }
 107     *d = '\0';
 108
 109     info("Received %zu bytes from graphite backend. Ignoring them. Sample: '%s'", buffer_strlen(b), sample);
 110     buffer_flush(b);
 111     return 0;
 112 }
 113
 114 static inline int process_opentsdb_response(BUFFER *b) {
 115     char sample[1024];
 116     const char *s = buffer_tostring(b);
 117     char *d = sample, *e = &sample[sizeof(sample) - 1];
 118
 119     for(; *s && d < e ;s++) {
 120         char c = *s;
 121         if(unlikely(!isprint(c))) c = ' ';
 122         *d++ = c;
 123     }
 124     *d = '\0';
 125
 126     info("Received %zu bytes from opentsdb backend. Ignoring them. Sample: '%s'", buffer_strlen(b), sample);
 127     buffer_flush(b);
 128     return 0;
 129 }
 130
 131 void *backends_main(void *ptr) {
 132     struct netdata_static_thread *static_thread = (struct netdata_static_thread *)ptr;
 133
 134     BUFFER *b = buffer_create(1), *response = buffer_create(1);
 135     int (*backend_request_formatter)(BUFFER *b, const char *prefix, RRDHOST *host, const char *hostname, RRDSET *st, RRDDIM *rd, time_t after, time_t before, uint32_t options) = NULL;
 136     int (*backend_response_checker)(BUFFER *b) = NULL;
 137
 138     info("BACKEND thread created with task id %d", gettid());
 139
 140     if(pthread_setcanceltype(PTHREAD_CANCEL_DEFERRED, NULL) != 0)
 141         error("Cannot set pthread cancel type to DEFERRED.");
 142
 143     if(pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, NULL) != 0)
 144         error("Cannot set pthread cancel state to ENABLE.");
 145
 146     // ------------------------------------------------------------------------
 147     // collect configuration options
 148
 149     struct timeval timeout = {
 150             .tv_sec = 0,
 151             .tv_usec = 0
 152     };
 153     int default_port = 0;
 154     int sock = -1;
 155     uint32_t options;
 156     int enabled = config_get_boolean("backend", "enabled", 0);
 157     const char *source = config_get("backend", "data source", "average");
 158     const char *type = config_get("backend", "type", "graphite");
 159     const char *destination = config_get("backend", "destination", "localhost");
 160     const char *prefix = config_get("backend", "prefix", "netdata");
 161     const char *hostname = config_get("backend", "hostname", localhost.hostname);
 162     int frequency = (int)config_get_number("backend", "update every", 10);
 163     int buffer_on_failures = (int)config_get_number("backend", "buffer on failures", 10);
 164     long timeoutms = config_get_number("backend", "timeout ms", frequency * 2 * 1000);
 165
 166     // ------------------------------------------------------------------------
 167     // validate configuration options
 168     // and prepare for sending data to our backend
 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(!strcmp(type, "graphite") || !strcmp(type, "graphite:plaintext")) {
 187         default_port = 2003;
 188         if(options == BACKEND_SOURCE_DATA_AS_COLLECTED)
 189             backend_request_formatter = format_dimension_collected_graphite_plaintext;
 190         else
 191             backend_request_formatter = format_dimension_stored_graphite_plaintext;
 192
 193         backend_response_checker = process_graphite_response;
 194     }
 195     else if(!strcmp(type, "opentsdb") || !strcmp(type, "opentsdb:telnet")) {
 196         default_port = 4242;
 197         if(options == BACKEND_SOURCE_DATA_AS_COLLECTED)
 198             backend_request_formatter = format_dimension_collected_opentsdb_telnet;
 199         else
 200             backend_request_formatter = format_dimension_stored_opentsdb_telnet;
 201
 202         backend_response_checker = process_opentsdb_response;
 203     }
 204     else {
 205         error("Unknown backend type '%s'", type);
 206         goto cleanup;
 207     }
 208
 209     if(backend_request_formatter == NULL || backend_response_checker == NULL) {
 210         error("backend is misconfigured - disabling it.");
 211         goto cleanup;
 212     }
 213
 214     if(timeoutms < 1) {
 215         error("BACKED invalid timeout %ld ms given. Assuming %d ms.", timeoutms, frequency * 2 * 1000);
 216         timeoutms = frequency * 2 * 1000;
 217     }
 218     timeout.tv_sec  = (timeoutms * 1000) / 1000000;
 219     timeout.tv_usec = (timeoutms * 1000) % 1000000;
 220
 221     // ------------------------------------------------------------------------
 222     // prepare the charts for monitoring the backend
 223
 224     struct rusage thread;
 225
 226     collected_number
 227             chart_buffered_metrics = 0,
 228             chart_lost_metrics = 0,
 229             chart_sent_metrics = 0,
 230             chart_buffered_bytes = 0,
 231             chart_received_bytes = 0,
 232             chart_sent_bytes = 0,
 233             chart_receptions = 0,
 234             chart_transmission_successes = 0,
 235             chart_transmission_failures = 0,
 236             chart_data_lost_events = 0,
 237             chart_lost_bytes = 0,
 238             chart_backend_reconnects = 0,
 239             chart_backend_latency = 0;
 240
 241     RRDSET *chart_metrics = rrdset_find("netdata.backend_metrics");
 242     if(!chart_metrics) {
 243         chart_metrics = rrdset_create("netdata", "backend_metrics", NULL, "backend", NULL, "Netdata Buffered Metrics", "metrics", 130600, frequency, RRDSET_TYPE_LINE);
 244         rrddim_add(chart_metrics, "buffered", NULL,  1, 1, RRDDIM_ABSOLUTE);
 245         rrddim_add(chart_metrics, "lost",     NULL,  1, 1, RRDDIM_ABSOLUTE);
 246         rrddim_add(chart_metrics, "sent",     NULL,  1, 1, RRDDIM_ABSOLUTE);
 247     }
 248
 249     RRDSET *chart_bytes = rrdset_find("netdata.backend_bytes");
 250     if(!chart_bytes) {
 251         chart_bytes = rrdset_create("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, RRDDIM_ABSOLUTE);
 253         rrddim_add(chart_bytes, "lost",     NULL, 1, 1024, RRDDIM_ABSOLUTE);
 254         rrddim_add(chart_bytes, "sent",     NULL, 1, 1024, RRDDIM_ABSOLUTE);
 255         rrddim_add(chart_bytes, "received", NULL, 1, 1024, RRDDIM_ABSOLUTE);
 256     }
 257
 258     RRDSET *chart_ops = rrdset_find("netdata.backend_ops");
 259     if(!chart_ops) {
 260         chart_ops = rrdset_create("netdata", "backend_ops", NULL, "backend", NULL, "Netdata Backend Operations", "operations", 130630, frequency, RRDSET_TYPE_LINE);
 261         rrddim_add(chart_ops, "write",     NULL, 1, 1, RRDDIM_ABSOLUTE);
 262         rrddim_add(chart_ops, "discard",   NULL, 1, 1, RRDDIM_ABSOLUTE);
 263         rrddim_add(chart_ops, "reconnect", NULL, 1, 1, RRDDIM_ABSOLUTE);
 264         rrddim_add(chart_ops, "failure",   NULL, 1, 1, RRDDIM_ABSOLUTE);
 265         rrddim_add(chart_ops, "read",      NULL, 1, 1, RRDDIM_ABSOLUTE);
 266     }
 267
 268     /*
 269      * this is misleading - we can only measure the time we need to send data
 270      * this time is not related to the time required for the data to travel to
 271      * the backend database and the time that server needed to process them
 272      *
 273      * issue #1432 and https://www.softlab.ntua.gr/facilities/documentation/unix/unix-socket-faq/unix-socket-faq-2.html
 274      *
 275     RRDSET *chart_latency = rrdset_find("netdata.backend_latency");
 276     if(!chart_latency) {
 277         chart_latency = rrdset_create("netdata", "backend_latency", NULL, "backend", NULL, "Netdata Backend Latency", "ms", 130620, frequency, RRDSET_TYPE_AREA);
 278         rrddim_add(chart_latency, "latency",   NULL,  1, 1000, RRDDIM_ABSOLUTE);
 279     }
 280     */
 281
 282     RRDSET *chart_rusage = rrdset_find("netdata.backend_thread_cpu");
 283     if(!chart_rusage) {
 284         chart_rusage = rrdset_create("netdata", "backend_thread_cpu", NULL, "backend", NULL, "NetData Backend Thread CPU usage", "milliseconds/s", 130630, frequency, RRDSET_TYPE_STACKED);
 285         rrddim_add(chart_rusage, "user",   NULL, 1, 1000, RRDDIM_INCREMENTAL);
 286         rrddim_add(chart_rusage, "system", NULL, 1, 1000, RRDDIM_INCREMENTAL);
 287     }
 288
 289     // ------------------------------------------------------------------------
 290     // prepare the backend main loop
 291
 292     info("BACKEND configured ('%s' on '%s' sending '%s' data, every %d seconds, as host '%s', with prefix '%s')", type, destination, source, frequency, hostname, prefix);
 293
 294     usec_t step_ut = frequency * USEC_PER_SEC;
 295     time_t after = now_realtime_sec();
 296     int failures = 0;
 297     heartbeat_t hb;
 298     heartbeat_init(&hb);
 299
 300     for(;;) {
 301         // ------------------------------------------------------------------------
 302         // Wait for the next iteration point.
 303         heartbeat_next(&hb, step_ut);
 304         time_t before = now_realtime_sec();
 305
 306         // ------------------------------------------------------------------------
 307         // add to the buffer the data we need to send to the backend
 308         RRDSET *st;
 309         int pthreadoldcancelstate;
 310
 311         if(unlikely(pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &pthreadoldcancelstate) != 0))
 312             error("Cannot set pthread cancel state to DISABLE.");
 313
 314         rrdhost_rdlock(&localhost);
 315         for(st = localhost.rrdset_root; st ;st = st->next) {
 316             pthread_rwlock_rdlock(&st->rwlock);
 317
 318             RRDDIM *rd;
 319             for(rd = st->dimensions; rd ;rd = rd->next) {
 320                 if(rd->last_collected_time.tv_sec >= after)
 321                     chart_buffered_metrics += backend_request_formatter(b, prefix, &localhost, hostname, st, rd, after, before, options);
 322             }
 323
 324             pthread_rwlock_unlock(&st->rwlock);
 325         }
 326         rrdhost_unlock(&localhost);
 327
 328         if(unlikely(pthread_setcancelstate(pthreadoldcancelstate, NULL) != 0))
 329             error("Cannot set pthread cancel state to RESTORE (%d).", pthreadoldcancelstate);
 330
 331         // ------------------------------------------------------------------------
 332
 333         chart_buffered_bytes = (collected_number)buffer_strlen(b);
 334
 335         // reset the monitoring chart counters
 336         chart_received_bytes =
 337         chart_sent_bytes =
 338         chart_sent_metrics =
 339         chart_lost_metrics =
 340         chart_transmission_successes =
 341         chart_transmission_failures =
 342         chart_data_lost_events =
 343         chart_lost_bytes =
 344         chart_backend_reconnects =
 345         chart_backend_latency = 0;
 346
 347         if(unlikely(netdata_exit)) break;
 348
 349         //fprintf(stderr, "\nBACKEND BEGIN:\n%s\nBACKEND END\n", buffer_tostring(b)); // FIXME
 350         //fprintf(stderr, "after = %lu, before = %lu\n", after, before);
 351
 352         // prepare for the next iteration
 353         // to add incrementally data to buffer
 354         after = before;
 355
 356         // ------------------------------------------------------------------------
 357         // if we are connected, receive a response, without blocking
 358
 359         if(likely(sock != -1)) {
 360             errno = 0;
 361
 362             // loop through to collect all data
 363             while(sock != -1 && errno != EWOULDBLOCK) {
 364                 buffer_need_bytes(response, 4096);
 365
 366                 ssize_t r = recv(sock, &response->buffer[response->len], response->size - response->len, MSG_DONTWAIT);
 367                 if(likely(r > 0)) {
 368                     // we received some data
 369                     response->len += r;
 370                     chart_received_bytes += r;
 371                     chart_receptions++;
 372                 }
 373                 else if(r == 0) {
 374                     error("Backend '%s' closed the socket", destination);
 375                     close(sock);
 376                     sock = -1;
 377                 }
 378                 else {
 379                     // failed to receive data
 380                     if(errno != EAGAIN && errno != EWOULDBLOCK) {
 381                         error("Cannot receive data from backend '%s'.", destination);
 382                     }
 383                 }
 384             }
 385
 386             // if we received data, process them
 387             if(buffer_strlen(response))
 388                 backend_response_checker(response);
 389         }
 390
 391         // ------------------------------------------------------------------------
 392         // if we are not connected, connect to a backend server
 393
 394         if(unlikely(sock == -1)) {
 395             usec_t start_ut = now_monotonic_usec();
 396             const char *s = destination;
 397             while(*s) {
 398                 const char *e = s;
 399
 400                 // skip separators, moving both s(tart) and e(nd)
 401                 while(isspace(*e) || *e == ',') s = ++e;
 402
 403                 // move e(nd) to the first separator
 404                 while(*e && !isspace(*e) && *e != ',') e++;
 405
 406                 // is there anything?
 407                 if(!*s || s == e) break;
 408
 409                 char buf[e - s + 1];
 410                 strncpyz(buf, s, e - s);
 411                 chart_backend_reconnects++;
 412                 sock = connect_to(buf, default_port, &timeout);
 413                 if(sock != -1) break;
 414                 s = e;
 415             }
 416             chart_backend_latency += now_monotonic_usec() - start_ut;
 417         }
 418
 419         if(unlikely(netdata_exit)) break;
 420
 421         // ------------------------------------------------------------------------
 422         // if we are connected, send our buffer to the backend server
 423
 424         if(likely(sock != -1)) {
 425             size_t len = buffer_strlen(b);
 426             usec_t start_ut = now_monotonic_usec();
 427             int flags = 0;
 428 #ifdef MSG_NOSIGNAL
 429             flags += MSG_NOSIGNAL;
 430 #endif
 431
 432             ssize_t written = send(sock, buffer_tostring(b), len, flags);
 433             chart_backend_latency += now_monotonic_usec() - start_ut;
 434             if(written != -1 && (size_t)written == len) {
 435                 // we sent the data successfully
 436                 chart_transmission_successes++;
 437                 chart_sent_bytes += written;
 438                 chart_sent_metrics = chart_buffered_metrics;
 439
 440                 // reset the failures count
 441                 failures = 0;
 442
 443                 // empty the buffer
 444                 buffer_flush(b);
 445             }
 446             else {
 447                 // oops! we couldn't send (all or some of the) data
 448                 error("Failed to write data to database backend '%s'. Willing to write %zu bytes, wrote %zd bytes. Will re-connect.", destination, len, written);
 449                 chart_transmission_failures++;
 450
 451                 if(written != -1)
 452                     chart_sent_bytes += written;
 453
 454                 // increment the counter we check for data loss
 455                 failures++;
 456
 457                 // close the socket - we will re-open it next time
 458                 close(sock);
 459                 sock = -1;
 460             }
 461         }
 462         else {
 463             error("Failed to update database backend '%s'", destination);
 464             chart_transmission_failures++;
 465
 466             // increment the counter we check for data loss
 467             failures++;
 468         }
 469
 470         if(failures > buffer_on_failures) {
 471             // too bad! we are going to lose data
 472             chart_lost_bytes += buffer_strlen(b);
 473             error("Reached %d backend failures. Flushing buffers to protect this host - this results in data loss on back-end server '%s'", failures, destination);
 474             buffer_flush(b);
 475             failures = 0;
 476             chart_data_lost_events++;
 477             chart_lost_metrics = chart_buffered_metrics;
 478         }
 479
 480         if(unlikely(netdata_exit)) break;
 481
 482         // ------------------------------------------------------------------------
 483         // update the monitoring charts
 484
 485         if(chart_ops->counter_done) rrdset_next(chart_ops);
 486         rrddim_set(chart_ops, "read",         chart_receptions);
 487         rrddim_set(chart_ops, "write",        chart_transmission_successes);
 488         rrddim_set(chart_ops, "discard",      chart_data_lost_events);
 489         rrddim_set(chart_ops, "failure",      chart_transmission_failures);
 490         rrddim_set(chart_ops, "reconnect",    chart_backend_reconnects);
 491         rrdset_done(chart_ops);
 492
 493         if(chart_metrics->counter_done) rrdset_next(chart_metrics);
 494         rrddim_set(chart_metrics, "buffered", chart_buffered_metrics);
 495         rrddim_set(chart_metrics, "lost",     chart_lost_metrics);
 496         rrddim_set(chart_metrics, "sent",     chart_sent_metrics);
 497         rrdset_done(chart_metrics);
 498
 499         if(chart_bytes->counter_done) rrdset_next(chart_bytes);
 500         rrddim_set(chart_bytes, "buffered",   chart_buffered_bytes);
 501         rrddim_set(chart_bytes, "lost",       chart_lost_bytes);
 502         rrddim_set(chart_bytes, "sent",       chart_sent_bytes);
 503         rrddim_set(chart_bytes, "received",   chart_received_bytes);
 504         rrdset_done(chart_bytes);
 505
 506         /*
 507         if(chart_latency->counter_done) rrdset_next(chart_latency);
 508         rrddim_set(chart_latency, "latency",  chart_backend_latency);
 509         rrdset_done(chart_latency);
 510         */
 511
 512         getrusage(RUSAGE_THREAD, &thread);
 513         if(chart_rusage->counter_done) rrdset_next(chart_rusage);
 514         rrddim_set(chart_rusage, "user",   thread.ru_utime.tv_sec * 1000000ULL + thread.ru_utime.tv_usec);
 515         rrddim_set(chart_rusage, "system", thread.ru_stime.tv_sec * 1000000ULL + thread.ru_stime.tv_usec);
 516         rrdset_done(chart_rusage);
 517
 518         if(likely(buffer_strlen(b) == 0))
 519             chart_buffered_metrics = 0;
 520
 521         if(unlikely(netdata_exit)) break;
 522     }
 523
 524 cleanup:
 525     if(sock != -1)
 526         close(sock);
 527
 528     buffer_free(b);
 529     buffer_free(response);
 530
 531     info("BACKEND thread exiting");
 532
 533     static_thread->enabled = 0;
 534     pthread_exit(NULL);
 535     return NULL;
 536 }