Lessons from High Velocity Logging Presented by Brad Lhotsky

Brad Lhotsky • Security and Systems Administrator
 at Craigslist
 • Previously: Infrastructure Monitoring & Security
 at Booking.com • Recovering • Perl Programmer • Linux Systems Admin • Network Security Specialist • PostgreSQL Administrator • ElasticSearch Janitor • DNS Voyeur • Starving Author • OSSEC Core Team Member https://github.com/reyjrar https://twitter.com/reyjrar

Amazing People Johnny Lopes, Cully Bennefield, Wouter Clarie, Willem Gerber, Lucian Grosu, Stefan Boronea, Wijnand Modderman-Lensta, Mickey Nasriachi, Mattia Barbon, Alexey Shatlovsky, Vadim Sukhanov, Daniel Ostermeier, Samit Badle, Tim Pickett, Nick Perez

Expectations ‣Lessons from Shipping Logs ‣Lessons from Parsing Logs ‣Lessons from Storing Logs ‣Lessons from Using Logs

Assumptions ‣Business Intelligence covered ‣Running Linux servers ‣Developed by Security Engineers Your Mileage May Vary

Shipping Logs at WEB SCALE!

Syslog Transport ‣Syslog is well-known transport ‣Syslog-ng installed on all servers ‣Don’t do that, use rsyslog ‣Remember syslog() syscall blocks

Syslog-ng source s_apache { file("/logs/access.log" follow_freq(1)); }; filter f_junk { not(match("favicon.ico")); }; destination d_remote { tcp("1.2.3.4" port(514)); }; log { source(s_apache); filter(f_junk); destination(d_remote); }; So Pretty! Any one see any bugs? ‣“favicon.ico” are silently discarded ‣Reads “/logs/access.log” once per second ‣Only reads log_fetch_limit() lines each read log_fetch_limit() default is 10

Configure rsyslog to: ‣Queue messages locally if the remote server dies ‣Sends all data over TCP ... ‣Encrypted ... ‣With peer validation

rsyslog # Configure GNU/TLS $DefaultNetstreamDriver gtls $DefaultNetstreamDriverCAFile /etc/pki/rsyslog/ca.pem $DefaultNetstreamDriverCertFile /etc/pki/rsyslog/client-cert.pem $DefaultNetstreamDriverKeyFile /etc/pki/rsyslog/client-key.pem $ActionQueueType LinkedList $ActionQueueFileName centralwork $ActionResumeRetryCount -1 $ActionQueueSaveOnShutdown on $ActionSendStreamDriverAuthMode x509/name $ActionSendStreamDriverPermittedPeer log.example.com $ActionSendStreamDriverMode 1 *.* @@1.2.3.4:514

rsyslog outputs ‣Forwarding, Files, Sockets ‣Programs, SNMP, Email ‣PostgreSQL, MySQL, Oracle, DBI ‣ElasticSearch, MongoDB, Kafka, HDFS Now that's webscale!

No content

Syslog Isn’t Web Scale 1. Apache Kafka 2. nanomsg 3. ZeroMQ, RabbitMQ, ActiveMQ, “${HOT}MQ” 4. Redis (just don’t tell anyone it’s not a queue) Syslog isn’t a Pub/Sub Queue

Syslog Isn’t Web Scale ‣ Apache Flume ‣ Facebook’s Scribe Syslog isn’t a Logging Pipeline

Syslog Isn’t Web Scale 1. Fluentd 2. LogStash 3. Hindsight Syslog isn’t a Logging Platform

Syslog Isn’t Web Scale Syslog is boring technology. We need more boring technology. Simplify our end points. Understand our Infrastructure.

Perl Isn’t Web Scale We did this in Perl with POE. It outperformed LogStash. 60,000 Messages/Second. Limited by ElasticSearch indexing speed.

Parsing Logs

Syslog Contents ‣ When did the log happen? ‣ Which server sent the log? (use FQDN!) ‣ What level was the log? ‣ What program generated the log? ‣ What was the content of the log?

Parse::Syslog::Line ‣ Order from chaos, logs to documents ‣ Fastest syslog parser on the CPAN ‣ Handle many variations on syslog formats ‣ Determine how much I can hate vendors Goals

Parse::Syslog::Line Turns this: <11>Jan 1 00:00:00 router.example.com dhcpd[3428]: DHCPINFORM from 172.16.2.137 via vlan3 Into this: --- content: DHCPINFORM from 172.16.2.137 via vlan3 date_str: 2015-01-01 00:00:00 datetime_str: 2015-01-01 00:00:00 domain: example.com epoch: 1420066800 facility: user facility_int: 8 host: router message: 'dhcpd[3428]: DHCPINFORM from 172.16.2.137 via vlan3' preamble: 11 priority: err priority_int: 3 program_name: dhcpd program_pid: 3428

Parse::Syslog::Line ‣ Vendors never comply with standards ‣ Always Const::Fast, never Readonly.pm ‣ Good regexes are fast ‣ Untested regexes are dangerously slow ‣ Sometimes, perldoc perlre lies Lessons https://xkcd.com/1171/

Regular Expressions? #!/usr/bin/env perl # use strict; use warnings; use Benchmark qw(cmpthese); my %RE = ( d => qr/\d+/ ); my $re = $RE{d}; my @tests = ; printf "Running regex benchmark with Perl %s\n", $^V; cmpthese(500_000, { 'with o' => sub { /$RE{d}/o for @tests }, 'without o' => sub { /$RE{d}/ for @tests }, 'variable' => sub { /$re/ for @tests }, 'static' => sub { /\d+/ for @tests, }, });

Regular Expressions? Running regex benchmark with Perl v5.24.0 Rate without o variable with o static without o 387597/s -- -7% -50% -54% variable 416667/s 7% -- -46% -51% with o 769231/s 98% 85% -- -9% static 847458/s 119% 103% 10% -- ‣ The %RE regex table is bad for performance ‣ perldoc perlre: o - pretend to optimize your code, but actually introduce bugs

Regular Expressions? for(@things) { if ( /$RE{IPv4}/o ) { print “Found IPv4!\n”; } } OK for(@things) { if ( /$RE{$type}/o ) { print “Found $type!\n”; } } NOT OK

Logs as Documents <134>Jun 2 07:49:17 bob-03.lhr4.prod.booking.com www.booking.com: 107.150.42.90 - - [02/Jun/2015:07:49:17 +0200] \"GET /plus/ search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*!50000And*/+(/*! 50000SeLECT*/+1+/*!50000frOM*/+(/*!50000SeLECT*/+/*! 50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*! 50000SeLECT*/+CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+` %23@__admin`+limit+0,1),1,62)))a+/*!50000fRom*/ +information_schema.tables+/*!50000gROUP*/+by+a)b)%23@`%5C'`+]=a HTTP/1.1\" 404 172775 0.368/0.368 \"http://www.booking.com/plus/ search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*!50000And*/+(/*! 50000SeLECT*/+1+/*!50000frOM*/+(/*!50000SeLECT*/+/*! 50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*! 50000SeLECT*/+CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+` %23@__admin`+limit+0,1),1,62)))a+/*!50000fRom*/ +information_schema.tables+/*!50000gROUP*/+by+a)b)%23@`%5C'`+]=a\" \"baidu; baiduspider;\" - 556D43DD00922CB5

Adding Context ‣ Using FQDN, get data center? ‣ Well known log formats can be tokenized ‣ Find embedded JSON and parse it Do’s Shouldn’ts ‣ Bad (read: untested) Regex ‣ HTTP API Calls ‣ DNS Lookups

Logs as Documents --- '@fields': crit: 404 dc: LHR4 domain: lhr4.prod.booking.com dst: www.booking.com facility: local0 file: /plus/search.php filetype: php hostname: bob-03 method: GET out_bytes: 172775 program: apache protocol: HTTP/1.1 referer: "http://www.booking.com/plus/search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*!50000And*/+(/*!50000SeLECT*/+1+/*!50000frOM*/+(/*! 50000SeLECT*/+/*!50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*!50000SeLECT*/+CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+` %23@__admin`+limit+0,1),1,62)))a+/*!50000fRom*/+information_schema.tables+/*!50000gROUP*/+by+a)b)%23@`%5C'`+]=a" resource: "/plus/search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*!50000And*/+(/*!50000SeLECT*/+1+/*!50000frOM*/+(/*!50000SeLECT*/+/*! 50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*!50000SeLECT*/+CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+`%23@__admin`+limit +0,1),1,62)))a+/*!50000fRom*/+information_schema.tables+/*!50000gROUP*/+by+a)b)%23@`%5C'`+]=a" response_ms: 368 service: web severity: info src_ip: 107.150.42.90 status: error upstream_ms: 368 useragent: baidu; baiduspider; zone: prod '@message': "www.booking.com: 107.150.42.90 - - [02/Jun/2015:07:49:17 +0200] \"GET /plus/search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*! 50000And*/+(/*!50000SeLECT*/+1+/*!50000frOM*/+(/*!50000SeLECT*/+/*!50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*!50000SeLECT*/ +CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+`%23@__admin`+limit+0,1),1,62)))a+/*!50000fRom*/+information_schema.tables+/*!50000gROUP*/ +by+a)b)%23@`%5C'`+]=a HTTP/1.1\" 404 172775 0.368/0.368 \"http://www.booking.com/plus/search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*! 50000And*/+(/*!50000SeLECT*/+1+/*!50000frOM*/+(/*!50000SeLECT*/+/*!50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*!50000SeLECT*/ +CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+`%23@__admin`+limit+0,1),1,62)))a+/*!50000fRom*/+information_schema.tables+/*!50000gROUP*/ +by+a)b)%23@`%5C'`+]=a\" \"baidu; baiduspider;\" - 556D43DD00922CB5" '@source_host': bob-03.lhr4.prod.booking.com '@timestamp': 2015-06-02T07:49:17+0200

Extract More Meta ‣ Parse URI’s for their parts ‣ Extract relevant fields from URI parameters ‣ Simple attack detection using common words ‣ GeoIP from a local DB

--- '@fields': attack_generic_score: 16 attack_score: 340 attack_sqli_score: 18 attack_triggers: '*/ | */ | */ | */ | */ | */ | */ | */ | /* | /* | / * | /* | /* | /* | /* | /* | concat | concat | rand | select | select | select' referer_domain: www.booking.com referer_file: /plus/search.php src_ip_country: US src_ip_location: '39.0997,-94.5786' Now That’s Meta?

Lessons Learned ‣ Test the performance of the system ‣ Work with the data you have ‣ Tokenize the data into a set of well defined fields

Storing Logs

Lessons from Breaking ElasticSearch ‣ 30 GB Heap Maximum ‣ Leave 33%+ Memory for FS Cache ‣ Google "Call Me Maybe ElasticSearch” ‣ Separate master and data nodes ‣ Name space indexes based on retention

ElasticSearch Shards ‣ Transaction Log ‣ Lucene Segments ‣ Optimize Condenses Segments

Optimized Shards ‣ Fewer segments ‣ Less Mapping ‣ Less Reducing ‣ Moar Web Scale ‣ Expensive operation

elasticsearch.yml # Disable the use of Memory-mapped files index.store.type: niofs # Indexing Settings for Writes index.refresh_interval: 30s index.merge.scheduler.max_thread_count: 1 index.translog.flush_threshold_size: 1gb # Field Data Cache indices.fielddata.cache.size: 40%

Index Layouts

Index Optimizations ‣ Split Indexes in functional namespaces ‣ Use Dynamic Mappings (Schemas) ‣ Never use auto_expand_replicas ‣ Fewer replicas yield faster writes

Document Optimizations ‣ Each doc uses ~8 bytes of memory ‣ Analyzing fields is expensive ‣ Agree on a “field set” for your documents ‣ CEE, CEF, XDAS, GELF, RELP, SDEE, etc ‣ Invent your own

Schemaless Data

Schemaless Data ‣ Can't optimize storage ‣ Can't optimize search ‣ But it's easy! ‣ Bad for performance ‣ Read ‣ Search ‣ Index / Create

Dates in ES { "date_template": { "match": "*_date", "mapping": { "type": "date", "format": "date_time||epoch_second", "ignore_malformed": true } } }, Matches things like: access_date, login_date

IP Addresses in ES { "ip_template": { "match": "*_ip", "mapping": { "type": "ip", "ignore_malformed": true, "fields": { "raw": { "type" : "string", "index": "not_analyzed" } } } } } Matches src_ip and creates src_ip.raw

Dates in ES { "int_template": { "match": "*_int", "mapping": { "type": "long", "ignore_malformed": true } } },

Geo-data in ES { "geo_template" : { "mapping" : { "type" : "geo_point" }, "match" : "*_geoip" } }, Rejects docs where *_geoip don't cast properly.

Strings { "string_fields" : { "mapping": { "type": "string", }, "match": "*", "match_mapping_type": "string" } }

Strings in ES 2.x { "string_fields" : { "mapping": { "type": "string", "index": "not_analyzed", "ignore_above": 256 }, "match": "*", "match_mapping_type": "string" } }

Strings in ES 5.x { "string_fields" : { "mapping": { "type": "keyword", "index": true, "ignore_above": 256 }, "match": "*", "match_mapping_type": "string" } } For k/v fields, ie: program:"apache"

Strings in ES 5.x { "message" : { "mapping": { "type": "text", "index": true } } } For bodies of text, like the whole log message

Strings in ES ‣Analyzers ‣ Store and Search ‣ Affect indexing of tokens ‣Tokenizers ‣ Boundaries between tokens ‣Token Filters ‣ Remove tokens ‣Character Filters ‣ Removing data from tokens https://xkcd.com/171/

Filter Example "insensitive_stop": { "type": "stop", "stopwords": "_english_", "ignore_case": true }, "min_length": { "type": "length", "min": 2 }

Filter Example "kill_numbers" : { "type": "pattern_replace", "pattern": "(\\d+)", "replace": "" }

Analyzer Example "cl_words_analyzer": { "tokenizer": "whitespace", "char_filter": "kill_numbers", "filter": [ "lowercase", "asciifolding", "insensitive_stop", "min_length", "kstem" ] }

Putting it Together message": { "type": "text", "index": true, "analyzer": "cl_words_analyzer", "norms": false, "fielddata": false, "fields": { "raw": { "type": "keyword", "index": "no", "doc_values": false, "norms": { "enabled": false }, "ignore_above": 1024 } } }

Object Storage "rule": { "type": "object", "enabled": true, "dynamic": false, "properties": { "name": { "type": "keyword", "ignore_above": 256 }, "id": { "type": "keyword", "ignore_above": 256 }, "tags": { "type": "keyword", "ignore_above": 256 }, "cve": { "type": "keyword", "ignore_above": 256 }, "info": { "type": "keyword", "ignore_above": 1024 }, "level": { "type": "long", "ignore_malformed": true } } },

Nested Object Storage "rules": { "type": "nested", "enabled": true, "dynamic": false, "properties": { "name": { "type": "keyword", "index": true }, "filter": { "type": "keyword", "index": true }, "condition": { "type": "keyword", "index": true }, "match": { "type": "keyword", "index": true }, "args": { "type": "object", "enabled": false } } }

No content

https://www.elastic.co/guide/en/elasticsearch/reference/5.0/ nested.html

Defensive Mapping index.mapping.total_fields.limit: 1000 index.mapping.depth.limit: 20 index.mapping.nested_fields.limit:50 ‣ You need to know your data ‣ Even when you do, safe guards are necessary ‣ Luckily, other people learned from this pain

Monitor and Maintain ‣ Delete old indexes ‣ Degrade replicas ‣ Close unused indexes ‣ Disable bloom filters ‣ Monitor key metrics

$ es-daily-index-maintenance.pl --base logstash \ --close --close-days 90 \ --delete --delete-days 180 \ --replicas --replicas-min 0 \ --replicas-max 2 \ --replicas-age 7 App::ElasticSearch::Utilities Maintain Indexes https://github.com/reyjrar/es-utils

App::ElasticSearch::Utilities Maintain Aliases # es-alias-manager.pl --config /etc/es-aliases.yml --- logstash: pattern: \*-logstash-{{DATE}} daily: logstash-{{DATE}} relative: alias: logstash-{{PERIOD}} periods: weekly: from: days: 7 to: days: 0 https://github.com/reyjrar/es-utils

Index Aliases

$ es-graphite-dynamic.pl App::ElasticSearch::Utilities Monitor https://github.com/reyjrar/es-utils

Using Logs

General Web Overview

Web Attacks Scanners

Sudoers Wall of Shame

Slow Pages

App::ElasticSearch::Utilities Search Stuff! = Querying Indexes: lhr4-access-2015.06.03,ams4-access-2015.06.03 @timestamp src_ip src_ip_country file 2015-06-03T11:39:27+0200 217.36.201.217 GB /B1D671CF- E532-4481-99AA-19F420D90332/netdefender/hui/ndhui.css 2015-06-03T11:39:26+0200 92.56.217.84 ES /hotel/es/ null.es.html # Search Parameters: # {"query_string":{"query":"dst:www.booking.com AND crit:404"}} # Displaying 3 of (CENSORED) in 0 seconds. # Indexes (2 of 4) searched: ams4-access-2015.06.03,lhr4-access-2015.06.03 https://github.com/reyjrar/es-utils $ es-search.pl --base access dst:www.booking.com and crit:404 \ --show src_ip,src_ip_country,file --size 2

Aggregate Stuff https://github.com/reyjrar/es-utils $ es-search.pl --base access --days 1 dst:www.booking.com \ --top src_ip --size 3 = Querying Indexes: ams4-access-2015.06.03,lhr4-access-2015.06.03 count src_ip (CENSORED) 66.249.92.71 (CENSORED) 66.249.92.59 (CENSORED) 66.249.92.65 # Search Parameters: # {"query_string":{"query":"dst:www.booking.com"}} # Displaying 3 of (CENSORED) in 5 seconds. # Indexes (2 of 2) searched: ams4-access-2015.06.03,lhr4-access-2015.06.03 # # Totals across batch # count src_ip (CENSORED) 66.249.92.71 (CENSORED) 66.249.92.59 (CENSORED) 66.249.92.65

Find Pages Viewed by Most Countries https://github.com/reyjrar/es-utils $ es-search.pl --base access --days 1 dst:www.booking.com \ --top file --by cardinality:src_ip_country --size 3 = Querying Indexes: lhr4-access-2015.06.03,ams4-access-2015.06.03 cardinality:src_ip_country count file 239 (CENSORED) / 236 (CENSORED) /rt_data/city_bookings 234 (CENSORED) /wishlist/get # Search Parameters: # {"query_string":{"query":"dst:www.booking.com"}} # Displaying 3 of (CENSORED) in 21 seconds. # Indexes (2 of 2) searched: ams4-access-2015.06.03,lhr4-access-2015.06.03

Pipeline Queries https://github.com/reyjrar/es-utils $ es-search.pl --base access --days 1 dst:www.booking.com \ --top src_ip --by sum:attack_score --size 3 \ --data-file top_attackers.dat $ es-search.pl --base access --days 1 dst:www.booking.com \ src_ip:top_attackers.dat[-1] --size 3\ --show attack_score,src_ip,crit,dst,method,resource \ --sort attack_score:desc

Pipeline Queries https://github.com/reyjrar/es-utils = Querying Indexes: lhr4-access-2015.06.03,ams4-access-2015.06.03 @timestamp attack_score src_ip crit dst method resource 2015-06-03T04:20:59+0200 340 107.150.42.90 404 www.booking.com GET /plus/ search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*!50000And*/+(/*!50000SeLECT*/+1+/*!50000frOM*/+(/*! 50000SeLECT*/+/*!50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*!50000SeLECT*/ +CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+`%23@__admin`+limit+0,1),1,62)))a+/*! 50000fRom*/+information_schema.tables+/*!50000gROUP*/+by+a)b)%23@`%5C'`+]=a 2015-06-03T00:50:43+0200 340 107.150.42.90 404 www.booking.com GET /plus/ search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*!50000And*/+(/*!50000SeLECT*/+1+/*!50000frOM*/+(/*! 50000SeLECT*/+/*!50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*!50000SeLECT*/ +CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+`%23@__admin`+limit+0,1),1,62)))a+/*! 50000fRom*/+information_schema.tables+/*!50000gROUP*/+by+a)b)%23@`%5C'`+]=a 2015-06-03T05:18:19+0200 340 107.150.42.90 404 www.booking.com GET /plus/ search.php?keyword=as&typeArr[111%3D@`%5C'`)+/*!50000And*/+(/*!50000SeLECT*/+1+/*!50000frOM*/+(/*! 50000SeLECT*/+/*!50000Count(*)*/,concat(floor(rand(0)*2),(substring((/*!50000SeLECT*/ +CONCAT(0x40,userid,0x7c,substring(pwd,4,16))+from+`%23@__admin`+limit+0,1),1,62)))a+/*! 50000fRom*/+information_schema.tables+/*!50000gROUP*/+by+a)b)%23@`%5C'`+]=a # Search Parameters: # {"terms":{"src_ip":["107.150.42.90","37.59.7.157","74.84.138.120"]}} # {"query_string":{"query":"dst:www.booking.com"}} # Displaying 3 of 793 in 0 seconds. # Indexes (2 of 2) searched: ams4-access-2015.06.03,lhr4-access-2015.06.03

Thank You! Presented by: Brad Lhotsky https://twitter.com/reyjrar https://github.com/reyjrar https://speakerdeck.com/reyjrar Made possible by the support of and viewers like you! craigslist