{"id":53531,"date":"2019-12-04T00:00:00","date_gmt":"2019-12-03T21:00:00","guid":{"rendered":"https:\/\/prohoster.info\/blog\/blog_prohoster\/ishhem-anomalii-i-predskazyvaem-sboi-s-pomoshhyu-nejrosetej"},"modified":"2020-02-18T14:01:26","modified_gmt":"2020-02-18T11:01:26","slug":"ishhem-anomalii-i-predskazyvaem-sboi-s-pomoshhyu-nejrosetej","status":"publish","type":"post","link":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/ishhem-anomalii-i-predskazyvaem-sboi-s-pomoshhyu-nejrosetej","title":{"rendered":"Hled\u00e1me anom\u00e1lie a p\u0159edpov\u00edd\u00e1me selh\u00e1n\u00ed pomoc\u00ed neuronov\u00fdch s\u00edt\u00ed","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

\"Hled\u00e1me<\/p>\n

<\/p>\n

Pr\u016fmyslov\u00fd v\u00fdvoj softwarov\u00fdch syst\u00e9m\u016f vy\u017eaduje velkou pozornost k odolnosti kone\u010dn\u00e9ho produktu, stejn\u011b jako rychlou reakci na selh\u00e1n\u00ed a poruchy, pokud k nim p\u0159ece jen dojde. Monitorov\u00e1n\u00ed samoz\u0159ejm\u011b pom\u00e1h\u00e1 reagovat na poruchy a selh\u00e1n\u00ed efektivn\u011bji a rychleji, ale nesta\u010d\u00ed to. Za prv\u00e9, velmi obt\u00ed\u017en\u00e9 je sledovat velk\u00fd po\u010det server\u016f \u2013 je pot\u0159eba zna\u010dn\u00e9 mno\u017estv\u00ed lid\u00ed. Za druh\u00e9, je d\u016fle\u017eit\u00e9 dob\u0159e porozum\u011bt tomu, jak je aplikace konstruov\u00e1na, abychom mohli p\u0159edpov\u011bd\u011bt jej\u00ed stav. To znamen\u00e1, \u017ee pot\u0159ebujeme mnoho lid\u00ed, kte\u0159\u00ed dob\u0159e rozum\u011bj\u00ed syst\u00e9m\u016fm, kter\u00e9 vyv\u00edj\u00edme, jejich ukazatel\u016fm a vlastnostem. P\u0159edpokl\u00e1dejme, \u017ee i kdyby se na\u0161lo dostate\u010dn\u00e9 mno\u017estv\u00ed lid\u00ed, ochotn\u00fdch se t\u00edmto zab\u00fdvat, vy\u017eaduje tak\u00e9 spoustu \u010dasu jejich \u0161kolen\u00ed.<\/p>\n

<\/p>\n

A co te\u010f? Na pomoc n\u00e1m sp\u011bch\u00e1 um\u011bl\u00e1 inteligence. Tato \u010dl\u00e1nek se bude zab\u00fdvat prediktivn\u00edm \u00fadr\u017eb\u00e1\u0159stv\u00edm<\/a><\/noindex> (predictive maintenance). Tento p\u0159\u00edstup aktivn\u011b z\u00edsk\u00e1v\u00e1 na popularit\u011b. Bylo naps\u00e1no velk\u00e9 mno\u017estv\u00ed \u010dl\u00e1nk\u016f, v\u010detn\u011b t\u011bch na Habr. Velk\u00e9 firmy ho intenzivn\u011b vyu\u017e\u00edvaj\u00ed k podpo\u0159e funk\u010dnosti sv\u00fdch server\u016f. Po prostudov\u00e1n\u00ed velk\u00e9ho mno\u017estv\u00ed \u010dl\u00e1nk\u016f jsme se rozhodli tento p\u0159\u00edstup vyzkou\u0161et. Co z toho vze\u0161lo? <\/p>\n

<\/a><\/noindex><\/p>\n

\u00davod<\/h2>\n

<\/p>\n

Vyvinut\u00fd softwarov\u00fd syst\u00e9m se d\u0159\u00edve nebo pozd\u011bji dost\u00e1v\u00e1 do provozu. Pro u\u017eivatele je d\u016fle\u017eit\u00e9, aby syst\u00e9m fungoval bez v\u00fdpadk\u016f. Pokud k mimo\u0159\u00e1dn\u00e9 situaci p\u0159ece jen dojde, m\u011bla by b\u00fdt odstran\u011bna s minim\u00e1ln\u00edmi zpo\u017ed\u011bn\u00edmi. <\/p>\n

<\/p>\n

Pro zjednodu\u0161en\u00ed technick\u00e9 podpory softwarov\u00e9ho syst\u00e9mu, zejm\u00e9na pokud je server\u016f mnoho, se obvykle pou\u017e\u00edvaj\u00ed monitorovac\u00ed programy, kter\u00e9 sb\u00edraj\u00ed metriky z b\u011b\u017e\u00edc\u00edho softwarov\u00e9ho syst\u00e9mu, umo\u017e\u0148uj\u00ed diagnostikovat jeho stav a pom\u00e1haj\u00ed ur\u010dit, co konkr\u00e9tn\u011b zp\u016fsobilo selh\u00e1n\u00ed. Tento proces se naz\u00fdv\u00e1 monitorov\u00e1n\u00ed softwarov\u00e9ho syst\u00e9mu.<\/p>\n

\n\"Hled\u00e1me<\/p>\n

Obr\u00e1zek 1. Rozhran\u00ed pro monitorov\u00e1n\u00ed grafana<\/em><\/p>\n

Metriky jsou r\u016fzn\u00e9 ukazatele programov\u00e9ho syst\u00e9mu, prost\u0159ed\u00ed jeho spu\u0161t\u011bn\u00ed nebo fyzick\u00e9ho v\u00fdpo\u010detn\u00edho stroje, pod kter\u00fdm je syst\u00e9m spu\u0161t\u011bn s \u010dasov\u00fdm raz\u00edtkem, kter\u00e9 ud\u00e1v\u00e1 okam\u017eik, kdy byly metriky z\u00edsk\u00e1ny. V statick\u00e9 anal\u00fdze jsou data metrik naz\u00fdv\u00e1na \u010dasov\u00fdmi \u0159ady. Pro sledov\u00e1n\u00ed stavu programov\u00e9ho syst\u00e9mu se metriky zobrazuj\u00ed ve form\u011b graf\u016f: na ose X je \u010das a na ose Y jsou hodnoty (obr\u00e1zek 1). Z funk\u010dn\u00edho programov\u00e9ho syst\u00e9mu lze zaznamen\u00e1vat n\u011bkolik tis\u00edc metrik (z ka\u017ed\u00e9ho uzlu). Ty tvo\u0159\u00ed prostor metrik (v\u00edcerozm\u011brn\u00e9 \u010dasov\u00e9 \u0159ady). <\/p>\n

<\/p>\n

Vzhledem k tomu, \u017ee u slo\u017eit\u00fdch programov\u00fdch syst\u00e9m\u016f je zaznamen\u00e1v\u00e1no velk\u00e9 mno\u017estv\u00ed metrik, ru\u010dn\u00ed monitorov\u00e1n\u00ed se st\u00e1v\u00e1 slo\u017eit\u00fdm \u00fakolem. Pro zmen\u0161en\u00ed objemu dat analyzovan\u00fdch spr\u00e1vcem obsahuj\u00ed monitorovac\u00ed n\u00e1stroje n\u00e1stroje pro automatick\u00e9 zji\u0161t\u011bn\u00ed mo\u017en\u00fdch probl\u00e9m\u016f. Nap\u0159\u00edklad lze nastavit spou\u0161t\u011b\u010d, kter\u00fd se aktivuje v p\u0159\u00edpad\u011b poklesu voln\u00e9ho m\u00edsta na disku pod ur\u010ditou \u00farove\u0148. Tak\u00e9 je mo\u017en\u00e9 automaticky diagnostikovat zastaven\u00ed serveru nebo kritick\u00e9 zpomalen\u00ed obsluhy. V praxi monitorovac\u00ed n\u00e1stroje dob\u0159e zvl\u00e1daj\u00ed detekci ji\u017e vznikl\u00fdch poruch nebo odhalen\u00ed jednoduch\u00fdch p\u0159\u00edznak\u016f budouc\u00edch poruch, ale celkov\u011b z\u016fst\u00e1v\u00e1 predikce mo\u017en\u00e9ho selh\u00e1n\u00ed pro n\u011b tvrd\u00fdm o\u0159\u00ed\u0161kem. Predikce prost\u0159ednictv\u00edm ru\u010dn\u00ed anal\u00fdzy metrik vy\u017eaduje zapojen\u00ed kvalifikovan\u00fdch specialist\u016f. Je neefektivn\u00ed. V\u011bt\u0161ina potenci\u00e1ln\u00edch poruch m\u016f\u017ee z\u016fstat nepov\u0161imnuta.<\/p>\n

<\/p>\n

V posledn\u00ed dob\u011b mezi velk\u00fdmi IT spole\u010dnostmi zab\u00fdvaj\u00edc\u00edmi se v\u00fdvojem softwaru roste obl\u00edbenost takzvan\u00e9ho prediktivn\u00edho \u00fadr\u017eby programov\u00fdch syst\u00e9m\u016f. Podstata tohoto p\u0159\u00edstupu spo\u010d\u00edv\u00e1 v hled\u00e1n\u00ed z\u00e1vad, kter\u00e9 vedou k degradaci syst\u00e9mu v ran\u00fdch f\u00e1z\u00edch, p\u0159ed jeho selh\u00e1n\u00edm, s vyu\u017eit\u00edm um\u011bl\u00e9 inteligence. Tento p\u0159\u00edstup zcela nevylu\u010duje ru\u010dn\u00ed monitorov\u00e1n\u00ed syst\u00e9mu. Slou\u017e\u00ed jako dopln\u011bk k procesu monitorov\u00e1n\u00ed jako celku. <\/p>\n

<\/p>\n

Hlavn\u00edm n\u00e1strojem realizace prediktivn\u00ed \u00fadr\u017eby je \u00faloha hled\u00e1n\u00ed anom\u00e1li\u00ed v \u010dasov\u00fdch \u0159ad\u00e1ch, proto\u017ee p\u0159i vzniku anom\u00e1lie<\/strong> v datech je velk\u00e1 pravd\u011bpodobnost, \u017ee po n\u011bjak\u00e9 dob\u011b dojde k chyb\u011b nebo selh\u00e1n\u00ed<\/strong>. Anom\u00e1lie je ur\u010dit\u00e9 odch\u00fdlen\u00ed od ukazatel\u016f softwarov\u00e9ho syst\u00e9mu, jako je identifikace degradace rychlosti vykon\u00e1v\u00e1n\u00ed dotazu ur\u010dit\u00e9ho typu nebo pokles pr\u016fm\u011brn\u00e9ho po\u010dtu obsluhovan\u00fdch po\u017eadavk\u016f p\u0159i konstantn\u00ed \u00farovni z\u00e1kaznick\u00fdch relac\u00ed.<\/p>\n

<\/p>\n

\u00dakol hled\u00e1n\u00ed anom\u00e1li\u00ed v softwarov\u00fdch syst\u00e9mech m\u00e1 svou speci\u00e1ln\u00ed povahu. V podstat\u011b je pot\u0159eba vyvinout nebo upravit existuj\u00edc\u00ed metody pro ka\u017ed\u00fd softwarov\u00fd syst\u00e9m, proto\u017ee hled\u00e1n\u00ed anom\u00e1li\u00ed velmi z\u00e1vis\u00ed na datech, ve kter\u00fdch je prov\u00e1d\u011bno, a data softwarov\u00fdch syst\u00e9m\u016f se velmi li\u0161\u00ed v z\u00e1vislosti na n\u00e1stroj\u00edch implementace syst\u00e9mu a\u017e po to, na jak\u00e9m v\u00fdpo\u010detn\u00edm stroji je spu\u0161t\u011bn.<\/p>\n

<\/p>\n

Metody hled\u00e1n\u00ed anom\u00e1li\u00ed p\u0159i p\u0159edpov\u00edd\u00e1n\u00ed selh\u00e1n\u00ed softwarov\u00fdch syst\u00e9m\u016f<\/h2>\n

<\/p>\n

P\u0159edev\u0161\u00edm je t\u0159eba \u0159\u00edci, \u017ee my\u0161lenka p\u0159edpov\u00edd\u00e1n\u00ed selh\u00e1n\u00ed byla inspirov\u00e1na \u010dl\u00e1nkem \u201eStrojov\u00e9 u\u010den\u00ed v IT monitorov\u00e1n\u00ed\u201c<\/a><\/noindex>. Pro ov\u011b\u0159en\u00ed \u00fa\u010dinnosti p\u0159\u00edstupu s automatick\u00fdm hled\u00e1n\u00edm anom\u00e1li\u00ed byl vybr\u00e1n softwarov\u00fd syst\u00e9m \u201eWeb-Konsolidace\u201c, kter\u00fd je jedn\u00edm z projekt\u016f spole\u010dnosti NPO \u201eKrista\u201c. Pro n\u011bj byl d\u0159\u00edve prov\u00e1d\u011bn ru\u010dn\u00ed monitoring podle z\u00edskan\u00fdch metrik. Proto\u017ee je syst\u00e9m pom\u011brn\u011b slo\u017eit\u00fd, je sledov\u00e1no velk\u00e9 mno\u017estv\u00ed metrik: ukazatele JVM (zat\u00ed\u017een\u00ed garbage collectoru), ukazatele OS, na kter\u00e9 se k\u00f3d vykon\u00e1v\u00e1 (virtu\u00e1ln\u00ed pam\u011b\u0165, % zat\u00ed\u017een\u00ed CPU OS), ukazatele s\u00edt\u011b (zat\u00ed\u017een\u00ed s\u00edt\u011b), samotn\u00e9ho serveru (zat\u00ed\u017een\u00ed CPU, pam\u011bti), metriky wildfly a vlastn\u00ed metriky aplikace pro v\u0161echny kritick\u00e9 subsyst\u00e9my. <\/p>\n

<\/p>\n

V\u0161echny metriky jsou z\u00edsk\u00e1v\u00e1ny ze syst\u00e9mu pomoc\u00ed graphite. P\u016fvodn\u011b byla pou\u017eita datab\u00e1ze whisper jako standardn\u00ed \u0159e\u0161en\u00ed pro grafanu, ale s r\u016fstem z\u00e1kaznick\u00e9 z\u00e1kladny graphite p\u0159estala st\u00edhat, vy\u010derp\u00e1vaj\u00edc propustnost diskov\u00e9ho subsyst\u00e9mu DC. Pot\u00e9 bylo rozhodnuto hledat efektivn\u011bj\u0161\u00ed \u0159e\u0161en\u00ed. V\u00fdb\u011br padl na graphite+clickhouse<\/a><\/noindex>, co\u017e umo\u017enilo v\u00fdrazn\u011b sn\u00ed\u017eit z\u00e1t\u011b\u017e na diskov\u00fd subsyst\u00e9m a p\u011bt a\u017e \u0161estkr\u00e1t zmen\u0161it zabran\u00fd diskov\u00fd prostor. N\u00ed\u017ee je uvedena sch\u00e9ma mechanismu sb\u011bru metrik pomoc\u00ed graphite+clickhouse (obr\u00e1zek 2).<\/p>\n

\n\"Hled\u00e1me
\n<\/p>\n

Obr\u00e1zek 2. Sch\u00e9ma sb\u011bru metrik<\/em><\/p>\n

Sch\u00e9ma je p\u0159evzata z intern\u00ed dokumentace. Zobrazuje v\u00fdm\u011bnu dat mezi grafana (u\u017eivatelsk\u00fdm rozhran\u00edm pro monitorov\u00e1n\u00ed, kter\u00e9 pou\u017e\u00edv\u00e1me) a graphite. Sn\u00edm\u00e1n\u00ed metrik z aplikace prov\u00e1d\u00ed samostatn\u00fd software \u2013 jmxtrans<\/a><\/noindex>. Ten je tak\u00e9 ukl\u00e1d\u00e1 do graphite.
\nSyst\u00e9m \u201eWeb-Konsolidace\u201c m\u00e1 \u0159adu vlastnost\u00ed, kter\u00e9 vytv\u00e1\u0159ej\u00ed probl\u00e9my pro p\u0159edpov\u00edd\u00e1n\u00ed selh\u00e1n\u00ed:<\/p>\n

<\/p>\n

    \n
  1. \u010dasto doch\u00e1z\u00ed ke zm\u011bn\u011b trendu. Pro tento softwarov\u00fd syst\u00e9m vych\u00e1zej\u00ed r\u016fzn\u00e9 verze. Ka\u017ed\u00e1 z nich p\u0159in\u00e1\u0161\u00ed zm\u011bny v softwarov\u00e9 \u010d\u00e1sti syst\u00e9mu. T\u00edmto zp\u016fsobem v\u00fdvoj\u00e1\u0159i p\u0159\u00edmo ovliv\u0148uj\u00ed metriky tohoto syst\u00e9mu a mohou vyvolat zm\u011bnu trendu; <\/li>\n
  2. povaha implementace a tak\u00e9 c\u00edle pou\u017eit\u00ed z\u00e1kazn\u00edky tohoto syst\u00e9mu \u010dasto vy\u017eaduj\u00ed anom\u00e1lie bez p\u0159edchoz\u00ed degradace; <\/li>\n
  3. procento anom\u00e1li\u00ed v\u016f\u010di cel\u00e9mu souboru dat je mal\u00e9 (< 5%); <\/li>\n
  4. mohou nast\u00e1vat mezery v z\u00edsk\u00e1v\u00e1n\u00ed ukazatel\u016f ze syst\u00e9mu. V n\u011bkter\u00fdch kr\u00e1tk\u00fdch intervalech \u010dasu se monitorovac\u00edmu syst\u00e9mu neda\u0159\u00ed z\u00edskat metriky. Nap\u0159\u00edklad, pokud je server p\u0159et\u00ed\u017een. Pro u\u010den\u00ed neuronov\u00e9 s\u00edt\u011b je to kritick\u00e9. Je nutn\u00e9 synteticky vyplnit mezery;<\/li>\n
  5. P\u0159\u00edpady s anom\u00e1liemi \u010dasto plat\u00ed pouze pro konkr\u00e9tn\u00ed \u010d\u00edslo\/m\u011bs\u00edc\/\u010das (sez\u00f3nnost). Tento syst\u00e9m m\u00e1 jasn\u011b stanoven\u00e1 pravidla pro pou\u017e\u00edv\u00e1n\u00ed jeho u\u017eivateli. Metriky jsou tedy relevantn\u00ed pouze pro konkr\u00e9tn\u00ed \u010das. Syst\u00e9m se m\u016f\u017ee pou\u017e\u00edvat nepravideln\u011b, pouze v n\u011bkter\u00fdch m\u011bs\u00edc\u00edch: selektivn\u011b podle roku. Vznikaj\u00ed situace, kdy stejn\u00e9 chov\u00e1n\u00ed metrik v jednom p\u0159\u00edpad\u011b m\u016f\u017ee v\u00e9st k selh\u00e1n\u00ed softwarov\u00e9ho syst\u00e9mu, zat\u00edmco ve druh\u00e9m ne.
    \nNejprve byly analyzov\u00e1ny metody detekce anom\u00e1li\u00ed v datech sledov\u00e1n\u00ed softwarov\u00fdch syst\u00e9m\u016f. V \u010dl\u00e1nc\u00edch na toto t\u00e9ma se p\u0159i mal\u00fdch procentech anom\u00e1li\u00ed ve vztahu k ostatn\u00edm datov\u00fdm soubor\u016fm nej\u010dast\u011bji doporu\u010duje pou\u017e\u00edt neuronov\u00e9 s\u00edt\u011b. <\/li>\n<\/ol>\n

    <\/p>\n

    Hlavn\u00ed logika pro vyhled\u00e1v\u00e1n\u00ed anom\u00e1li\u00ed pomoc\u00ed dat neuronov\u00fdch s\u00edt\u00ed je zn\u00e1zorn\u011bna na obr\u00e1zku 3:<\/p>\n

    \n\"Hled\u00e1me
    \n<\/p>\n

    Obr\u00e1zek 3. Vyhled\u00e1v\u00e1n\u00ed anom\u00e1li\u00ed pomoc\u00ed neuronov\u00e9 s\u00edt\u011b<\/em><\/p>\n

    Na v\u00fdsledku p\u0159edpov\u011bdi nebo obnoven\u00ed okna aktu\u00e1ln\u00edho toku metrik se vypo\u010d\u00edt\u00e1v\u00e1 odchylka od z\u00edskan\u00e9ho s funguj\u00edc\u00edm softwarov\u00fdm syst\u00e9mem. V p\u0159\u00edpad\u011b velk\u00e9ho rozd\u00edlu mezi z\u00edskan\u00fdmi metrikami od softwarov\u00e9ho syst\u00e9mu a neuronov\u00e9 s\u00edt\u011b lze usoudit na anom\u00e1lii aktu\u00e1ln\u00edho segmentu dat. Vznik\u00e1 n\u00e1sleduj\u00edc\u00ed \u0159ada probl\u00e9m\u016f pro vyu\u017eit\u00ed neuronov\u00fdch s\u00edt\u00ed:<\/p>\n

    <\/p>\n

      \n
    1. pro spr\u00e1vnou pr\u00e1ci v streamovac\u00edm re\u017eimu mus\u00ed data pro tr\u00e9nink model\u016f neuronov\u00fdch s\u00edt\u00ed zahrnovat pouze \u201enorm\u00e1ln\u00ed\u201c data; <\/li>\n
    2. je nutn\u00e9 m\u00edt aktu\u00e1ln\u00ed model pro spr\u00e1vn\u00e9 odhalen\u00ed. Zm\u011bna trendu a sez\u00f3nnosti v metrik\u00e1ch m\u016f\u017ee vyvolat velk\u00e9 mno\u017estv\u00ed fale\u0161n\u00fdch poplach\u016f modelu. Pro jeho aktualizaci je t\u0159eba p\u0159esn\u011b ur\u010dit okam\u017eik, kdy je model zastaral\u00fd. Pokud se model aktualizuje pozd\u011bji nebo d\u0159\u00edve, pravd\u011bpodobn\u011b dojde k velk\u00e9mu po\u010dtu fale\u0161n\u00fdch poplach\u016f.
      \nNelze tak\u00e9 zapomenout na hled\u00e1n\u00ed a prevenci \u010dast\u00e9ho v\u00fdskytu fale\u0161n\u00fdch poplach\u016f. P\u0159edpokl\u00e1d\u00e1 se, \u017ee tyto budou nej\u010dast\u011bji vznikat v nestandardn\u00edch situac\u00edch. Mohou v\u0161ak b\u00fdt tak\u00e9 d\u016fsledkem chyby neuronov\u00e9 s\u00edt\u011b z d\u016fvodu nedostate\u010dn\u00e9ho tr\u00e9nov\u00e1n\u00ed. Je nutn\u00e9 minimalizovat po\u010det fale\u0161n\u00fdch poplach\u016f modelu. Jinak fale\u0161n\u00e9 p\u0159edpov\u011bdi zaberou hodn\u011b \u010dasu administr\u00e1tora, kter\u00fd je ur\u010den k ov\u011b\u0159en\u00ed syst\u00e9mu. D\u0159\u00edve nebo pozd\u011bji to skon\u010d\u00ed t\u00edm, \u017ee administr\u00e1tor p\u0159estane reagovat na \"paranoidu\" monitorovac\u00edho syst\u00e9mu.<\/li>\n<\/ol>\n

      <\/p>\n

      Rekurentn\u00ed neuronov\u00e1 s\u00ed\u0165<\/h2>\n

      <\/p>\n

      Pro odhalov\u00e1n\u00ed anom\u00e1li\u00ed v \u010dasov\u00fdch \u0159ad\u00e1ch lze pou\u017e\u00edt rekurentn\u00ed neuronovou s\u00ed\u0165 <\/a><\/noindex>s pam\u011bt\u00ed LSTM. Probl\u00e9m je pouze v tom, \u017ee ji lze pou\u017e\u00edt pouze pro p\u0159edpov\u00eddan\u00e9 \u010dasov\u00e9 \u0159ady. V na\u0161em p\u0159\u00edpad\u011b nejsou v\u0161echny metriky p\u0159edv\u00eddateln\u00e9. Pokus aplikovat RNN LSTM na \u010dasovou \u0159adu je zn\u00e1zorn\u011bn na obr\u00e1zku 4.<\/p>\n

      \n\"Hled\u00e1me
      \n<\/p>\n

      Obr\u00e1zek 4. P\u0159\u00edklad \u010dinnosti rekurentn\u00ed neuronov\u00e9 s\u00edt\u011b s bu\u0148kami pam\u011bti LSTM<\/em><\/p>\n

      Jak je vid\u011bt na obr\u00e1zku 4, RNN LSTM se poda\u0159ilo zvl\u00e1dnout detekci anom\u00e1li\u00ed v tomto \u010dasov\u00e9m \u00faseku. Tam, kde m\u00e1 v\u00fdsledek vysokou chybu p\u0159edpov\u011bdi (pr\u016fm\u011brn\u00e1 chyba), do\u0161lo skute\u010dn\u011b k anom\u00e1lii podle ukazatel\u016f. Pou\u017eit\u00ed pouze jedn\u00e9 RNN LSTM bude zjevn\u011b nedosta\u010duj\u00edc\u00ed, proto\u017ee je aplikovateln\u00e9 na mal\u00e9 mno\u017estv\u00ed metrik. M\u016f\u017eeme pou\u017e\u00edt jako dopl\u0148kovou metodu k hled\u00e1n\u00ed anom\u00e1li\u00ed. <\/p>\n

      <\/p>\n

      Autoenkod\u00e9r pro predikci poruch<\/h2>\n

      <\/p>\n

      Autoenkod\u00e9r<\/a><\/noindex> \u2013 v podstat\u011b um\u011bl\u00e1 neuronov\u00e1 s\u00ed\u0165. Vstupn\u00ed vrstva \u2013 encoder, v\u00fdstupn\u00ed vrstva \u2013 decoder. Nev\u00fdhodou v\u0161ech neuronov\u00fdch s\u00edt\u00ed tohoto typu je \u0161patn\u00e1 lokalizace anom\u00e1li\u00ed. Byla vybr\u00e1na architektura synchronn\u00edho autoenkod\u00e9ru.<\/p>\n

      \n\"Hled\u00e1me
      \n<\/p>\n

      Obr\u00e1zek 5. P\u0159\u00edklad \u010dinnosti autoenkod\u00e9ru<\/em><\/p>\n

      Autoenkod\u00e9ry se tr\u00e9nuj\u00ed na norm\u00e1ln\u00edch datech a pot\u00e9 hledaj\u00ed n\u011bco anom\u00e1ln\u00edho v datech pod\u00e1van\u00fdch do modelu. P\u0159esn\u011b to, co je pro tento \u00fakol pot\u0159eba. Z\u016fst\u00e1v\u00e1 jen vybrat, kter\u00fd autoenkod\u00e9r se hod\u00ed pro tento \u00fakol. Architektonicky nejjednodu\u0161\u0161\u00ed forma autoenkod\u00e9ru je p\u0159\u00edm\u00e1, jednosm\u011brn\u00e1 neuronov\u00e1 s\u00ed\u0165, kter\u00e1 velmi p\u0159ipom\u00edn\u00e1 v\u00edcestup\u0148ov\u00fd perceptron<\/a><\/noindex> (multilayer perceptron, MLP), s vstupn\u00ed, v\u00fdstupn\u00ed vrstvou a jedn\u00edm nebo n\u011bkolika skryt\u00fdmi vrstvami, kter\u00e9 je propojuj\u00ed.
      \nRozd\u00edly mezi autoenkod\u00e9ry a MLP spo\u010d\u00edvaj\u00ed v\u0161ak v tom, \u017ee v automatick\u00e9m kod\u00e9ru m\u00e1 v\u00fdstupn\u00ed vrstva stejn\u00fd po\u010det uzl\u016f jako vstupn\u00ed, a \u017ee m\u00edsto u\u010den\u00ed se predikovat c\u00edlovou hodnotu Y p\u0159i dan\u00e9m vstupu X se autoenkod\u00e9r u\u010d\u00ed rekonstruovat sv\u00e9 vlastn\u00ed X. Proto jsou autoenkod\u00e9ry modely pro neregulovan\u00e9 u\u010den\u00ed. <\/p>\n

      <\/p>\n

      \u00daloha autoenkod\u00e9ru spo\u010d\u00edv\u00e1 v nalezen\u00ed \u010dasov\u00fdch index\u016f r0 \u2026 rn, odpov\u00eddaj\u00edc\u00edch anom\u00e1ln\u00edm element\u016fm ve vstupn\u00edm vektoru X. Tento efekt je dosa\u017een hled\u00e1n\u00edm kvadratick\u00e9 chyby.<\/p>\n

      \n\"Hled\u00e1me <\/p>\n

      Obr\u00e1zek 6. Synchronn\u00ed autoenkod\u00e9r<\/em><\/p>\n

      Pro autoenkod\u00e9r byla vybr\u00e1na synchronn\u00ed architektura<\/a><\/noindex>. Jej\u00ed v\u00fdhody: mo\u017enost pou\u017eit\u00ed streamov\u00e9ho zpracov\u00e1n\u00ed a relativn\u011b men\u0161\u00ed po\u010det parametr\u016f neuronov\u00e9 s\u00edt\u011b ve srovn\u00e1n\u00ed s jin\u00fdmi architekturami.<\/p>\n

      <\/p>\n

      Mechanismus minimalizace fale\u0161n\u00fdch poplach\u016f<\/h2>\n

      <\/p>\n

      Vzhledem k r\u016fzn\u00fdm mimo\u0159\u00e1dn\u00fdm situac\u00edm a mo\u017enosti nedostate\u010dn\u00e9ho \u0161kolen\u00ed neuronov\u00e9 s\u00edt\u011b bylo rozhodnuto o nezbytnosti vyvinout mechanismus pro minimalizaci fale\u0161n\u00fdch poplach\u016f pro vyv\u00edjen\u00fd model detekce anom\u00e1li\u00ed. Tento mechanismus se zakl\u00e1d\u00e1 na z\u00e1klad\u011b \u0161ablon, kterou klasifikuje administr\u00e1tor. <\/p>\n

      <\/p>\n

      Algoritmus dynamick\u00e9 transformace \u010dasov\u00e9 osy<\/a><\/noindex> (algoritmus DTW, z anglick\u00e9ho dynamic time warping) umo\u017e\u0148uje nal\u00e9zt optim\u00e1ln\u00ed shodu mezi \u010dasov\u00fdmi sekvencemi. Poprv\u00e9 byl pou\u017eit v rozpozn\u00e1v\u00e1n\u00ed \u0159e\u010di: slou\u017eil k ur\u010den\u00ed, jak dva \u0159e\u010dov\u00e9 sign\u00e1ly p\u0159edstavuj\u00ed stejnou p\u016fvodn\u011b vyslovenou fr\u00e1zi. V pozd\u011bj\u0161\u00edch letech byl tak\u00e9 aplikov\u00e1n v dal\u0161\u00edch oborech.<\/p>\n

      <\/p>\n

      Hlavn\u00edm principem minimalizace fale\u0161n\u00fdch poplach\u016f je shroma\u017e\u010fov\u00e1n\u00ed datab\u00e1ze vzor\u016f pomoc\u00ed oper\u00e1tora, kter\u00fd klasifikuje podez\u0159el\u00e9 p\u0159\u00edpady, kter\u00e9 byly detekov\u00e1ny pomoc\u00ed neuronov\u00fdch s\u00edt\u00ed. N\u00e1sledn\u011b prob\u00edh\u00e1 porovn\u00e1n\u00ed zkategorizovan\u00e9ho vzoru s p\u0159\u00edpadem, kter\u00fd syst\u00e9m detekoval, a posuzuje se, zda se jedn\u00e1 o fale\u0161n\u00fd poplach nebo situaci vedouc\u00ed k selh\u00e1n\u00ed. Pro porovn\u00e1n\u00ed dvou \u010dasov\u00fdch \u0159ad se pou\u017e\u00edv\u00e1 algoritmus DTW. Hlavn\u00edm n\u00e1strojem minimalizace z\u016fst\u00e1v\u00e1 klasifikace. P\u0159edpokl\u00e1d\u00e1 se, \u017ee po shrom\u00e1\u017ed\u011bn\u00ed velk\u00e9ho mno\u017estv\u00ed vzorov\u00fdch p\u0159\u00edpad\u016f syst\u00e9m p\u0159estane \u010dasto vy\u017eadovat z\u00e1sah oper\u00e1tora z d\u016fvodu podobnosti v\u011bt\u0161iny p\u0159\u00edpad\u016f a vzniku podobn\u00fdch.<\/p>\n

      <\/p>\n

      Na z\u00e1klad\u011b v\u00fd\u0161e uveden\u00fdch metod neuronov\u00fdch s\u00edt\u00ed byla vytvo\u0159ena experiment\u00e1ln\u00ed aplikace pro predikci selh\u00e1n\u00ed syst\u00e9mu \u201eWeb-Konsolidace\u201c. C\u00edlem t\u00e9to aplikace bylo, za pou\u017eit\u00ed existuj\u00edc\u00edho archivu dat monitorov\u00e1n\u00ed a informac\u00ed o ji\u017e prob\u011bhl\u00fdch selh\u00e1n\u00edch, posoudit \u00fa\u010dinnost t\u00e9to metody pro na\u0161e softwarov\u00e9 syst\u00e9my. Sch\u00e9ma pr\u00e1ce aplikace je zn\u00e1zorn\u011bno n\u00ed\u017ee, na obr\u00e1zku 7.<\/p>\n

      \n\"Hled\u00e1me <\/p>\n

      Obr\u00e1zek 7. Sch\u00e9ma predikce selh\u00e1n\u00ed na z\u00e1klad\u011b anal\u00fdzy prostotu metrik<\/em><\/p>\n

      Na sch\u00e9matu lze vy\u010dlenit dv\u011b hlavn\u00ed oblasti: vyhled\u00e1v\u00e1n\u00ed anom\u00e1ln\u00edch \u010dasov\u00fdch \u00fasek\u016f v toku monitorovac\u00edch dat (metriku) a mechanismus minimalizace fale\u0161n\u00fdch poplach\u016f. Pozn\u00e1mka: pro experiment\u00e1ln\u00ed \u00fa\u010dely jsou data z\u00edsk\u00e1v\u00e1na prost\u0159ednictv\u00edm JDBC p\u0159ipojen\u00ed z datab\u00e1ze, do kter\u00e9 jsou ukl\u00e1d\u00e1na s pomoc\u00ed graphite.
      \nN\u00e1sleduje rozhran\u00ed v\u00fdsledn\u00e9ho syst\u00e9mu monitorov\u00e1n\u00ed (obr\u00e1zek 8).<\/p>\n

      \n\"Hled\u00e1me<\/p>\n

      Obr\u00e1zek 8. Rozhran\u00ed experiment\u00e1ln\u00edho syst\u00e9mu monitorov\u00e1n\u00ed<\/em><\/p>\n

      Na rozhran\u00ed se zobrazuje procento anom\u00e1lnosti podle z\u00edskan\u00fdch metrik. V na\u0161em p\u0159\u00edpad\u011b je v\u00fdstup simulov\u00e1n. Ji\u017e m\u00e1me v\u0161echna data za n\u011bkolik t\u00fddn\u016f a postupn\u011b je nahr\u00e1v\u00e1me pro ov\u011b\u0159en\u00ed p\u0159\u00edpadu anom\u00e1lie vedouc\u00ed k selh\u00e1n\u00ed. V doln\u00edm stavov\u00e9m \u0159\u00e1dku se zobrazuje celkov\u00e9 procento anom\u00e1lnosti dat v dan\u00e9m okam\u017eiku, kter\u00e9 je ur\u010deno pomoc\u00ed autoenkod\u00e9ru. Tak\u00e9 pro p\u0159edpov\u00eddan\u00e9 metriky se zobrazuje zvl\u00e1\u0161tn\u00ed procento, kter\u00e9 vypo\u010d\u00edt\u00e1v\u00e1 RNN LSTM.<\/p>\n

      <\/p>\n

      P\u0159\u00edklad detekce anom\u00e1lie ukazatel\u016f CPU pomoc\u00ed neuronov\u00e9 s\u00edt\u011b RNN LSTM (obr\u00e1zek 9).<\/p>\n

      \n\"Hled\u00e1me<\/p>\n

      Obr\u00e1zek 9. Detekce RNN LSTM<\/em><\/p>\n

      Pom\u011brn\u011b jednoduch\u00fd p\u0159\u00edpad, v podstat\u011b b\u011b\u017en\u00fd v\u00fdkyv, kter\u00fd v\u0161ak vedl k selh\u00e1n\u00ed syst\u00e9mu, byl \u00fasp\u011b\u0161n\u011b identifikov\u00e1n pomoc\u00ed RNN LSTM. Ukazatel anom\u00e1lnosti v tomto \u010dasov\u00e9m \u00faseku se rovn\u00e1 85 \u2013 95 %, v\u0161e nad 80 % (pr\u00e1h ur\u010den\u00fd experiment\u00e1ln\u011b) se pova\u017euje za anom\u00e1lii.
      \nP\u0159\u00edklad detekce anom\u00e1lie, kdy se syst\u00e9m neda\u0159ilo spustit po aktualizaci. Tuto situaci detekuje autoenkod\u00e9r (obr\u00e1zek 10).<\/p>\n

      \n\"Hled\u00e1me<\/p>\n

      Obr\u00e1zek 10. P\u0159\u00edklad detekce pomoc\u00ed autoenkod\u00e9ru<\/em><\/p>\n

      Jak je vid\u011bt na obr\u00e1zku, PermGen z\u016fstal na jedn\u00e9 \u00farovni. Autoenkod\u00e9r to pova\u017eoval za zvl\u00e1\u0161tn\u00ed, proto\u017ee d\u0159\u00edve nic podobn\u00e9ho nevid\u011bl. Anom\u00e1lnost zde setrv\u00e1v\u00e1 na 100 % do doby, ne\u017e se syst\u00e9m vr\u00e1t\u00ed do funk\u010dn\u00edho stavu. Anom\u00e1lnost se zobrazuje ve v\u0161ech metrik\u00e1ch. Jak ji\u017e bylo uvedeno, autoenkod\u00e9r nedok\u00e1\u017ee lokalizovat anom\u00e1lie. Oper\u00e1tor m\u00e1 tuto funkci plnit v t\u011bchto situac\u00edch.<\/p>\n

      <\/p>\n

      Z\u00e1v\u011br<\/h2>\n

      <\/p>\n

      PC \"Web-Konsolidace\" se vyv\u00edj\u00ed ji\u017e n\u011bkolik let. Syst\u00e9m je v pom\u011brn\u011b stabiln\u00edm stavu a po\u010det registrovan\u00fdch incident\u016f nen\u00ed p\u0159\u00edli\u0161 velk\u00fd. I p\u0159esto se poda\u0159ilo nal\u00e9zt anom\u00e1lie, kter\u00e9 vedly k selh\u00e1n\u00ed 5 \u2013 10 minut p\u0159ed jeho vznikem. V \u0159ad\u011b p\u0159\u00edpad\u016f by ozn\u00e1men\u00ed o selh\u00e1n\u00ed v\u010das pomohlo u\u0161et\u0159it pl\u00e1novan\u00fd \u010das vy\u010dlen\u011bn\u00fd na prov\u00e1d\u011bn\u00ed \"opravn\u00fdch\" prac\u00ed.<\/p>\n

      <\/p>\n

      Z experiment\u016f, kter\u00e9 se poda\u0159ilo prov\u00e9st, je je\u0161t\u011b brzy na definitivn\u00ed z\u00e1v\u011bry. V sou\u010dasnosti jsou v\u00fdsledky rozporupln\u00e9. Na jedn\u00e9 stran\u011b je patrn\u00e9, \u017ee algoritmy zalo\u017een\u00e9 na neur\u00e1ln\u00edch s\u00edt\u00edch jsou schopny nach\u00e1zet \"u\u017eite\u010dn\u00e9\" anom\u00e1lie. Na druh\u00e9 stran\u011b z\u016fst\u00e1v\u00e1 vysok\u00e9 procento fale\u0161n\u00fdch zji\u0161t\u011bn\u00ed a ne v\u0161echny anom\u00e1lie, kter\u00e9 kvalifikovan\u00fd specialista identifikuje, se da\u0159\u00ed neur\u00e1ln\u00ed s\u00edti zachytit. Mezi nev\u00fdhody pat\u0159\u00ed i to, \u017ee sou\u010dasn\u00e1 neur\u00e1ln\u00ed s\u00ed\u0165 vy\u017eaduje u\u010den\u00ed pod dohledem pro norm\u00e1ln\u00ed fungov\u00e1n\u00ed.<\/p>\n

      <\/p>\n

      Pro dal\u0161\u00ed rozvoj syst\u00e9mu predikce selh\u00e1n\u00ed a jeho dolad\u011bn\u00ed do uspokojiv\u00e9ho stavu lze zv\u00e1\u017eit n\u011bkolik cest. To zahrnuje podrobn\u011bj\u0161\u00ed anal\u00fdzu p\u0159\u00edpad\u016f s anom\u00e1liemi, kter\u00e9 vedou k selh\u00e1n\u00ed, co\u017e by vedlo k dopln\u011bn\u00ed seznamu d\u016fle\u017eit\u00fdch metrik, kter\u00e9 maj\u00ed velk\u00fd vliv na stav syst\u00e9mu, a k vylou\u010den\u00ed t\u011bch, kter\u00e9 nemaj\u00ed vliv. Pokud se budeme ub\u00edrat t\u00edmto sm\u011brem, mohli bychom tak\u00e9 u\u010dinit pokusy o specializaci algoritm\u016f konkr\u00e9tn\u011b pro na\u0161e p\u0159\u00edpady anom\u00e1li\u00ed, kter\u00e9 vedou k selh\u00e1n\u00ed. Existuje i jin\u00e1 cesta. To je zdokonalen\u00ed architektur neur\u00e1ln\u00edch s\u00edt\u00ed a t\u00edm zv\u00fd\u0161en\u00ed p\u0159esnosti detekce p\u0159i zkr\u00e1cen\u00ed \u010dasu u\u010den\u00ed.<\/p>\n

      <\/p>\n

      Vyjad\u0159uji pod\u011bkov\u00e1n\u00ed koleg\u016fm, kte\u0159\u00ed mi pomohli s psan\u00edm a udr\u017eov\u00e1n\u00edm aktu\u00e1lnosti tohoto \u010dl\u00e1nku: Viktoru Verbitsk\u00e9mu<\/a><\/noindex> a Sergeji Finogenovovi.<\/p>\n

      Zdroj: habr.com<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"

      \u041f\u0440\u043e\u043c\u044b\u0448\u043b\u0435\u043d\u043d\u0430\u044f \u0440\u0430\u0437\u0440\u0430\u0431\u043e\u0442\u043a\u0430 \u043f\u0440\u043e\u0433\u0440\u0430\u043c\u043c\u043d\u044b\u0445 \u0441\u0438\u0441\u0442\u0435\u043c \u0442\u0440\u0435\u0431\u0443\u0435\u0442 \u0431\u043e\u043b\u044c\u0448\u043e\u0433\u043e \u0432\u043d\u0438\u043c\u0430\u043d\u0438\u044f \u043a \u043e\u0442\u043a\u0430\u0437\u043e\u0443\u0441\u0442\u043e\u0439\u0447\u0438\u0432\u043e\u0441\u0442\u0438 \u043a\u043e\u043d\u0435\u0447\u043d\u043e\u0433\u043e \u043f\u0440\u043e\u0434\u0443\u043a\u0442\u0430, \u0430 \u0442\u0430\u043a\u0436\u0435 \u0431\u044b\u0441\u0442\u0440\u043e\u0433\u043e \u0440\u0435\u0430\u0433\u0438\u0440\u043e\u0432\u0430\u043d\u0438\u044f \u043d\u0430 \u043e\u0442\u043a\u0430\u0437\u044b \u0438 \u0441\u0431\u043e\u0438, \u0435\u0441\u043b\u0438 \u043e\u043d\u0438 \u0432\u0441\u0435-\u0442\u0430\u043a\u0438 \u0441\u043b\u0443\u0447\u0430\u044e\u0442\u0441\u044f. \u041c\u043e\u043d\u0438\u0442\u043e\u0440\u0438\u043d\u0433, \u043a\u043e\u043d\u0435\u0447\u043d\u043e \u0436\u0435, \u043f\u043e\u043c\u043e\u0433\u0430\u0435\u0442 \u0440\u0435\u0430\u0433\u0438\u0440\u043e\u0432\u0430\u0442\u044c \u043d\u0430 \u043e\u0442\u043a\u0430\u0437\u044b \u0438 \u0441\u0431\u043e\u0438 \u044d\u0444\u0444\u0435\u043a\u0442\u0438\u0432\u043d\u0435\u0435 \u0438 \u0431\u044b\u0441\u0442\u0440\u0435\u0435, \u043d\u043e \u043d\u0435\u0434\u043e\u0441\u0442\u0430\u0442\u043e\u0447\u043d\u043e. \u0412\u043e-\u043f\u0435\u0440\u0432\u044b\u0445, \u043e\u0447\u0435\u043d\u044c \u0441\u043b\u043e\u0436\u043d\u043e \u0443\u0441\u043b\u0435\u0434\u0438\u0442\u044c \u0437\u0430 \u0431\u043e\u043b\u044c\u0448\u0438\u043c \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e\u043c \u0441\u0435\u0440\u0432\u0435\u0440\u043e\u0432 \u2013 \u043d\u0435\u043e\u0431\u0445\u043e\u0434\u0438\u043c\u043e \u0431\u043e\u043b\u044c\u0448\u043e\u0435 \u043a\u043e\u043b\u0438\u0447\u0435\u0441\u0442\u0432\u043e \u043b\u044e\u0434\u0435\u0439. \u0412\u043e-\u0432\u0442\u043e\u0440\u044b\u0445, \u043d\u0443\u0436\u043d\u043e \u0445\u043e\u0440\u043e\u0448\u043e \u043f\u043e\u043d\u0438\u043c\u0430\u0442\u044c, \u043a\u0430\u043a […]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[688],"tags":[],"class_list":["post-53531","post","type-post","status-publish","format-standard","hentry","category-administrirovanie"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\n","aioseo_head_json":{"title":"\ud83e\udd47Hled\u00e1me anom\u00e1lie a predikujeme selh\u00e1n\u00ed s pomoc\u00ed neur\u00e1ln\u00edch s\u00edt\u00ed | ProHoster","description":"Pr\u016fmyslov\u00fd v\u00fdvoj.","canonical_url":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/ishhem-anomalii-i-predskazyvaem-sboi-s-pomoshhyu-nejrosetej","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"cs_CZ","og:site_name":"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b","og:type":"article","og:title":"\ud83e\udd47\u0418\u0449\u0435\u043c \u0430\u043d\u043e\u043c\u0430\u043b\u0438\u0438 \u0438 \u043f\u0440\u0435\u0434\u0441\u043a\u0430\u0437\u044b\u0432\u0430\u0435\u043c \u0441\u0431\u043e\u0438 \u0441 \u043f\u043e\u043c\u043e\u0449\u044c\u044e \u043d\u0435\u0439\u0440\u043e\u0441\u0435\u0442\u0435\u0439 | ProHoster","og:description":"\u041f\u0440\u043e\u043c\u044b\u0448\u043b\u0435\u043d\u043d\u0430\u044f \u0440\u0430\u0437\u0440\u0430\u0431\u043e\u0442\u043a\u0430.","og:url":"https:\/\/prohoster.info\/cs\/blog\/administrirovanie\/ishhem-anomalii-i-predskazyvaem-sboi-s-pomoshhyu-nejrosetej","og:image":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:secure_url":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:width":350,"og:image:height":350,"article:published_time":"2019-12-03T21:00:00+00:00","article:modified_time":"2020-02-18T11:01:26+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster","article:author":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"53531","title":null,"description":null,"keywords":null,"keyphrases":null,"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":null,"og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"","isEnabled":true},"graphs":[]},"schema_type":null,"schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":null,"robots_max_videopreview":null,"robots_max_imagepreview":"large","priority":null,"frequency":null,"local_seo":null,"seo_analyzer_scan_date":"2026-01-24 07:46:24","breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2021-02-28 20:23:55","updated":"2026-01-24 07:46:24","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/53531","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/comments?post=53531"}],"version-history":[{"count":0,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/posts\/53531\/revisions"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/media?parent=53531"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/categories?post=53531"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/cs\/wp-json\/wp\/v2\/tags?post=53531"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}