Browsed by
Category: Tech

Telegram’s infrastructure and outages. Some updates.

Telegram’s infrastructure and outages. Some updates.

This post is meant as an (ongoing) sequence of updates to the previous one about Telegram’s outages in March and April 2018. Please read it here first.

Last updated: April 30th, 6:00 AM UTC

UPDATE 1

With the help of a friend (and his own HowIsResolved), we managed to confirm that for most open resolvers worldwide (25k+ tested) api.telegram.org is showing up as 149.154.167.220. Only outliers seem to be China (resolving as of now as 174.37.154.236) and Russia (85.142.29.248).

UPDATE 2

During my analysis this morning I created a new Telegram App, and the (only) suggested MTProto (the Telegram protocol) server was 149.154.167.50. This falls into the IP range analysed above, and seems to be solely located in Amsterdam.

Kipters was so kind to review and help me notice that this server is not used for real “data” communication, but just for a “discovery” API call (help.getConfig method) which will return the list of servers that will have to be used for sending messages. We are currently still in process of comparing ranges received across the world, but in the best case scenario (ie: they are spread over multiple geographic locations) this would mean that there is still a single point of failure in the hardcoded “directory” server.

UPDATE 3

What I found in the previous note was “too weird to be true”, so I went ahead and kept digging into TDLib and the official Desktop and Android Apps, to confirm wether they were bootstrapping a session beginning from a single MTProto endpoint or not.

Fortunately, turns out this is not the case (relevant snippets for TDLib, DesktopAndroid Apps): both of them contain, hardcoded, in addition to endpoints in the range 149.154.167.0/24 (Amsterdam, AS62041), endpoints in 149.154.175.0/24 (Miami, AS59930) and 149.154.171.0/24 (Singapore, AS62014).

Sounds like we should look into different reasons why many users worldwide outside of EMEA had issues today (or wait for an official, detailed post mortem if it will ever come): there are many, from broken dependencies to weird cases of mis-routing.

Some areas are left to explore (feel free to share your ideas if you have any): why third party apps don’t have access to the whole list of “initial” MTProto endpoints, and are pushed to use only a single, non redundant one? Why the main website and api.telegram.org (mainly used for bots I think) are based off a single location?

UPDATE 4

Telegram Web (https://web.telegram.org/) seems to be single-homed in Amsterdam too. As I haven’t had the opportunity to test during the outage, I don’t know whether it has been failed over somewhere else or not.

UPDATE 5

According to the official documentation, users (registered by phone number) are located off a single datacenter, picked at signup time based on geographical proximity: “During the process of working with the API, user information is accumulated in the DC with which the user is associated. This is the reason a user cannot be associated with a different DC by means of the client.

They are only moved if they keep connecting from a remote location for a prolonged period of time (ie: you permanently relocate to another continent): this might explain why there seem to be no failover scenario and 12+ hours outages are happening.

(Thanks to adjustableneutralism from Reddit for flagging)

Telegram is down (again): a deep look at their infrastructure.

Telegram is down (again): a deep look at their infrastructure.

I’ve been a strong Telegram advocate since its launch in 2013, mainly because of the advanced features and technical state of the art compared to competitors – as a consequence, I’ve been looking very closely at their infrastructure for the last few years.

The two large scale outages that recently hit their users and the sequence of events following them made me ask some questions around their platform. For most of them I have only found additional question marks rather than answers, but here it is what I have so far.

Let’s start from the outages: in case you missed that, on March 29th and April 29th this year, Telegram went down in their Amsterdam datacenter due to a power failure, causing disruptions, according to their official communications, to users in EMEA, MENA, Russia and CIS.

Zooming in on the latter: it’s still ongoing at time of writing this article (8:30AM UTC), and is showing up with clients unable to connect to the platform and both https://www.telegram.org/ (website) and https://api.telegram.org/ (api endpoint) failing with an HTTP error code 500.

Let’s start with the items that, to me, don’t add up: first and foremost, the outage. In case of “massive power outage” in the Amsterdam area, I would expect to see a traffic drop in AMS-IX, the largest Internet Exchange in the region, but there is none (it should be showing around 01 AM):

There are indeed reports of an outage that affected Amsterdam (below the one from Schiphol Airport), but no (public) reports of consequent large datacenter failures.

Who’s involved in running large scale platforms will be surprised by at least two things here: the fact that they are serving an huge geographical area from a single datacenter and their inability to reactively reroute traffic to the other locations they are operating, even in case of extended outage (no DR plans?).

A quick search on Twitter shows that even if the official communication states the issue is only affecting the EMEA region, users from Canada, US, Australia, Japan and other countries are facing it as well.

I used Host-Tracker to have a deeper look into this: an HTTP check to Telegram’s API endpoint and their website fails with an HTTP 500 error from every location across the world:

I went ahead and began digging to find out more about their infrastructure, network and the other locations they are running from.

And here comes the second huge question mark: the infrastructure.

A bunch of DNS lookups across the main endpoints show they are always resolving to the same v4 and v6 IPs, in a way that doesn’t look related to the source location of my queries.

They look to be announced by AS62041 (owned by Telegram LLP): this kind of DNS scheme made me think they were running an anycast based network, so next logical step has been analysing latencies from multiple locations.

Turns out, latency is averaging 20/30ms from EMEA, 100/150ms from AMER, and 250/300ms from APAC: as if from all of those countries you were being routed to the Amsterdam datacenter.

What I’m seeing in terms of latency is confirmed by analysing reverse lookups of routers found in the different paths to Telegram: in my trace from Australia the last visible hop is et3-1-2.amster1.ams.seabone.net (notice that “ams”), most of the traces from US are landing on xcr1.att.cw.net (195.2.1.14) which 1 millisecond away from my lab in Amsterdam and a couple of samples from US and Canada are running all the way up to ae-2-3201.ear3.Amsterdam1.Level3.net, which is self-explaining.

Important to highlight, there are no outliers: I couldn’t find a single example of very low latency from APAC / AMER, that would have proved the existence of a local point of presence. A summary of my tests in the table below:

To get the full picture, I decided to dig into AS62041 main upstream carriers (CW AS1273, TI Sparkle AS6762, Level3 AS3356) and see how they were handing over internet traffic to Telegram.

Turns out, CW is always preferring the path to xcr1.att.cw.net/195.2.1.14 (tested from some locations across the world), our little router-friend in Amsterdam. TI Sparkle always lands on amster1.ams.seabone.net and Level3 only has paths to ear3.Amsterdam1 (tested from Asia and US). Level3’s BGP communities are interesting: routes are tagged as “Europe Backbone” and “Level3_Customer Netherlands Amsterdam”:

Telegram is also peering with Hurricane Electric (AS6939): their routers in US, JP, AU have a next hop of ams-ix-gw.telegram.org/80.249.209.69 for 149.154.164.0/22. That hop seems to be Telegram’s AMS-IX facing router, and the IP is definitely part of AMS-IX:

 

As said in the opening, there are definitely more questions than answers in the article. It’s as if there was no Telegram infrastructure outside Amsterdam, and over there it was running in a single datacenter. This would explain why users across the world are seeing an outage that should only affect EMEA and close areas, and why Telegram is not taking steps to reroute users to another datacenter/location during the failure in AMS.

Am I missing something very obvious? Please let me know!

UPDATE: With the help of some friends and random people, I found out more details. Find them (with -ongoing- updates) in the dedicated post.

This is your sysadmin speaking: please expect some turbulence.

This is your sysadmin speaking: please expect some turbulence.

A few months back I blogged about my HP DL320 Gen8’s (in)compatibility with the outside world, and someone suggested me to solve the problem by replacing the P420i RAID controller with an LSI-something which would ensure wider flexibility.

Others were suggesting to replace (again) the hard drives instead, and someone was even pushing to swap this “hobby” with something healthier and go cloud instead*.

For the first time in my life I decided to listen to friends, so I replaced the RAID controller with an LSI 9300i HBA (I’m using mdraid anyway)…

…well, not really: I also replaced the chassis, motherboard, CPU, RAM banks, fans, PSUs and drive caddies.

Meet “ZA Rev2″**:

This is how it evolved:

  • HP -> Supermicro (yay!)
  • Xeon E3-1240 v2 -> Xeon E3-1240 v6
  • 4×8 GB DDR3 RAM -> 2×16 GB DDR4 RAM (2 slots free for future upgrades)
  • HP P420i -> LSI-9300i
  • 2x SSD Samsung 850 EVO 250 GB -> no change
  • 2x HGST SATA 7.2k 1 TB -> no change

D-Day for replacement is April 18th (taking a day off from my job to go and do the same things, just for hobby, feels really weird, yes), with a 6 AM wake up call, flight to AMS, 8/10 hours to do everything and a flight back to LON (LTN to be precise, because I didn’t double check before hitting “Buy”).

Now to the sad part: there is no (easy) way to just move the drives to the new server and have everything working, so I have to reinstall it from the ground up. This means my stuff (including this blog, because loose-coupling is a thing but I decided to run its DB and NFS from another country… …for some reason) will be down (or badly broken) during that time window and possibly longer, depending how much I manage to do while I’m onsite.

The timing couldn’t be better for a clean start, as in the last few months I had been considering the option to move away (escape) from Proxmox (which, as an example, is so flexible that its management port number is hardcoded everywhere and can’t be changed) to something else, most likely oVirt or OpenNebula. Haven’t taken a decision yet, but I’ve really fallen in love with the latter: it’s perfect for the cloud-native minds and runs on Debian, whereas oVirt would force me to move to the RPM side of the world.

Deeply apologise in advance for my rants on Twitter while I try to accomplish this mission. Stay tuned.

Giorgio

 

* I.AM.100%.CLOUD. There are two things you can’t (yet) do in the cloud: physical backup of your assets that live in the cloud and testing stuff which requires VT extensions. This is what I’m doing here: ZA is my bare-metal lab.

** this is not ZA Rev2. It was supposed to be, but it came in with a faulty backplane so I pushed for it to be entirely replaced. I don’t have a picture of the new one with me at the time of writing but… yeah, it looks exactly the same (with better cable management).

Eventi straordinari e siti istituzionali: un rapporto (ancora) tormentato.

Eventi straordinari e siti istituzionali: un rapporto (ancora) tormentato.

Anni fa ho scritto questo articolo (in un momento di frustrazione causata dalla puntuale indisponibilità dei siti istituzionali nei momenti di loro maggiore utilità), nella speranza quantomeno di aprire una linea di dialogo. Ero stato fortunato e questa si era aperta, ma il tutto era stato impacchettato e rispedito al mittente senza troppi complimenti.

Il problema in breve: sono molti i siti informativi, soprattutto in ambito Pubblica Amministrazione, “inutili” e poco visitati per il 99.9% del tempo, che però diventano critici in momenti di particolare interesse. Immaginate ad esempio il censimento della popolazione: ha cadenza decennale e dura due mesi. Durante questa finestra di tempo ogni cittadino userà l’apposito servizio online, ovviamente aspettandosi che tutto funzioni a dovere.

Altro esempio è il portale del Ministero dell’Istruzione: basso carico per gran parte dell’anno, ma quando vengono annunciate le commissioni di maturità, deve essere funzionante, pronto e scattante. Pensate poi al sito dove vengono pubblicati i risultati delle elezioni: utilizzato ogni quattro o cinque anni, diventa il più visitato d’Italia durante le poche ore di scrutinio.

Internet oggi è la fonte primaria di informazione per molte persone: è un dato di fatto che non si può ignorare, ed è necessario dare adeguata importanza alle piattaforme che contribuiscono a questa informazione.

Ne parlavo nel 2011, perchè è stato l’anno in cui i tre servizi sopracitati hanno mancato il loro obiettivo primario: quando servivano, non funzionavano. Se ne era parlato, soprattutto tra gli addetti ai lavori: ci eravamo arrabbiati, ma qualcuno aveva commentato che le soluzioni al problema (che spaziano da questioni molto tecniche come lo sharding dei database e l’elasticità delle infrastrutture a questioni più di buon senso, come una corretta previsione dei carichi) erano molto distanti dal mondo dei “comuni mortali”, e ancor di più dal settore pubblico.

Un punto di vista secondo me contestabile, ma quasi sicuramente con un fondo di verità: al tempo il concetto di “cloud” esisteva da pochi anni, e alcuni vendor dubitavano ancora delle sue potenzialità.

Sembra di parlare della preistoria.

(per non dimenticare: il load balancing manuale delle Elezioni 2011)

Adesso siamo nel 2017: sono passati sei anni dal mio articolo e come alcuni continuano a ripetere, “cloud is the new normal”. Il cloud è la nuova normalità, tutti lo usano, lo scetticismo, se mai c’è stato, è sparito: il tempo ha ormai provato che è una nuova e rivoluzionaria tecnologia e non solo un trend temporaneo o una pazzia di un singolo vendor.

In questi anni, nella nostra PA, sarà cambiato qualcosa?

Alcuni segnali fanno ben sperare: Eligendo ad esempio, il portale delle Elezioni, è esposto tramite una CDN (ma non supporta HTTPS). Altri fanno invece perdere la speranza appena guadagnata: questo mese si è tenuto il Referendum per l’Autonomia della Lombardia – serve che vi dica in che stato era il sito ufficiale durante gli scrutini? Timeout.

Le soluzioni a questo tipo di problemi sono ormai ben conosciute e consolidate: caching estremo, utilizzo di CDN, sfruttamento di infrastrutture scalabili, etc. I costi sono molto bassi e granulari: con una architettura ben studiata, si possono servire tutte le richieste senza sprecare un euro. Fa in un certo senso pensare il fatto che in certi ambienti siano ancora presenti e gravi problemi che l’industria ha risolto già da tempo, come quello dei picchi di carico.

Quali sono quindi i fattori limitanti, quindi?

Non stento a credere ci sia una scarsa comprensione del tema e della sua importanza ai “piani alti” di ogni ente: solo di recente siamo riusciti a mettere insieme una community di sviluppatori e un “team digitale” (composto da professionisti di veramente alto rango) volto a svecchiare il “sistema Italia”.

L’iniziativa sta già portando i suoi primi frutti, ma si tratta di un team per ora piccolo molto focalizzato sullo sviluppo e non sulle operations/mantenimento: il passo per il cambiamento della mentalità generale è ancora lungo. Non è difficile immaginare come una scarsa comprensione del tema porti molto velocemente alla mancanza di interesse e di risorse dedicate – con conseguente frustrazione di quelli che sono i “piani inferiori”.

Un secondo fattore spesso portato (o meglio, trascinato) in gioco è la scarsità di infrastrutture: se questo poteva essere vero una volta, oggi, con l’affermazione delle tecnologie cloud e del concetto di “on demand”, questo smette di essere un punto bloccante. Le infrastrutture ci sono, basta sfruttarle.

Ultimo, ma non per importanza, il discorso “competenze”: non stento a credere come molti fanno notare che sia difficile reclutare personale adatto e che chi si occupa oggi di sistemi nella PA abbia ben altre responsabilità e quindi ben altre basi. Ritengo però non si possa ignorare il fatto che al giorno d’oggi il concetto di “as a service” (servizi managed se volete chiamarli con un nome forse più familiare) rimuova buona parte di questo problema, e che l’immensa offerta di training e relativa facilità di sperimentazione renda estremamente facile la coltivazione delle skills mancanti.

Può servire tempo, ma da qualche parte bisognerà pur partire. Molti IT manager e sistemisti sono lì fuori pronti, a fare il passo: hanno solo bisogno di essere ispirati.

Ispiriamoli, no?

Me and HP: a “bare-metal” odyssey

Me and HP: a “bare-metal” odyssey

If you follow me on social media you’ve almost certainly heard of the issues I’m facing with the HP DL320 Gen8 I bought a few months back to replace my NAS and some test machines.

In term of diagnosing and solving this problem HP’s tech support has been useless so far, so in the last few weeks I’ve been digging deeper and deeper into this, and here are my findings (in logical, and not chronological, order).

Let’s start from scratch, for the benefit of who has not been following this from the very beginning: I’ve installed, tested and shipped the machine with the main drives only (Samsung 850 EVO SSD), as the capacity ones I wanted to use (SATA 2.5″, 1TB, 7200rpm) turned out not being easy to find on the market.

When I was finally able to buy 2+1 drives of the exact HGST model I was after, I screwed them to their caddies and shipped them to the colocation: when they confirmed the drives had been placed into the server, I rebooted it and configured them in a mirrored mdraid array.

Then I noticed that power consumption had gone up from 0.3 to 0.5 Amps:

The raid (re)build was still ongoing and CPU usage was high, so I just ignored this, even if even during previous spikes I had never seen such an high power consumption. To mi surprise, the morning after power usage was still 0.5A, even if the rebuild had finished hours before and load average was back to 0.0something.

With no evidence of something being wrong with the system itself, I blamed the drives (HGST HTS721010A9E630) and started researching for someone else facing the same issue with them. Nothing came out, as expected, and got confirmation from some docs that the power usage to be expected was way lower than what I was seeing.

By chance, I found some threads on the HP forums mentioning situations where non-genuine hard drives were causing “high noise”. Being unable to check the noise by myself without travelling to the colocation, I went ahead and had a look at the fans speed in my iLO, to realise all of them were running at 100%: at that stage I didn’t knew the pre-upgrade reading (now I do: 19%), but while testing it at home (in a way less controlled and warmer environment than the datacenter) I had never seen anything above 30%.

At this stage, I had finally found the cause for that huge power usage: extremely high fan speed. It was now time to try and explain the latter. First thing I checked, of course, were temperatures around the system: everything was good according to the iLO, no alarms nor criticals (not even warnings) and SMART readings were fine, with 20/21C on every drive. Nothing was explaining why the DL320 was trying so hard to cool itself down.

Then I found this article, where David described the same problem and found the perfect name for this phenomenon: Thermal Runway. Based on his description, looks like I’ve been very lucky, as other HP ProLiant servers are even shutting themselves down due to wrong temperature readings. Needless to say, my hard drives P/N were in its list of known bad ones.

Scraping the IPMI details, I found the sensor who was causing this whole thing: “05-HD Max”, which was at 58C. I’ve researched its details, and looks like it’s not a physical sensor, but rather an average of all of the SMART readings. With the temps for my four drives being around 22/23C max according to SMART, there was no way their average could have been 58C. Making things worse, this sensor has an hardcoded, non editable warning threshold at 60C.

With no clue on what to do next, I tried asking HGST if there was a firmware upgrade available (the DL320 G8 is on latest version of everything), but after 15 days, a number of emails and multiple levels of escalation they didn’t even manage to understand what I was asking for, so I decided to give up with them.

At this stage, with all the details I was able to gather I logged a support case to HP, and at the same time bought two new Seagate HDDs (ST500LM021-1KJ15), just to learn, after trying them, that they cause the same problem.

After a very honest first answer where HP’s tech support told me that the system was speeding up FANs as the drives were not recognised as HP genuine, they changed their mind and started pretending the 58C reading was real, and my drives were really running so hot.

I was lost again, and started wondering what did prehistoric people do before the cloud came, when they had this kind of hardware issues. Their first step was probably to go in front of the broken server, so I jumped on a plane and did the same.

(a picture of MY-ZA while undergoing surgery)

First thing, I was able to confirm the 58C reading was definitely wrong (as expected, anyway, but I was looking for a proof to show HP), and SMART was right: drives were super-cold, even if extracted while running. Moreover that sensor was jumping from 24C to 58C in 2/3 seconds after placing them in, which is rather hard (just think about the thermal shock).

Second, I tried to put the drives in different positions (and on a different port of the P420i RAID controller), and the issue was still there.

As last resort, I connected them to the onboard B120i HBA, and the system started working properly. Sensor 05 back to normal, drives running ok, etc. Not a good solution tough, as I’ve paid for the P420i + cache and under no circumstance I will do without it.

Fortunately, while upgrading my iLO4 to firmware 2.55, I noticed that after resetting it sensor 05 was temporarily disappearing, until the next operating system reboot. With this sensor disappearing, everything goes back to normal: fans to 30%, consumption to 0.3A, my bank account not at risk anymore.

sensor 05 has disappeared: 03, 04, … 06.

So, even if not particularly good looking and clean, I had found a solution: resetting the iLO. I went ahead and installed freeipmi, then made sure “bmc-device –cold-reset” is run 30 seconds after the system boots.

I’m still holding some kind of hope in HP support: I asked them to provide me with a way to permanently disable that sensor or raise its threshold, at my risk (read: voiding warranty).

It’s hard to describe how frustrated I am with both with HP servers, policies and support: not being able to test all existing parts and so having some “genuine” and some non genuine ones is okay, but artificially messing up a temperature reading to increase power consumption (and thus costs) and force their customers not using parts from 3rd parties can only be defined with a word: sabotage.

Giorgio

%d bloggers like this: