Tuesday, October 6, 2015

Reflections on (Humans) Trusting (Humans') Trust

One thing I've learned from consulting is that you should trust a competent person's feels about a technology, even if they can't immediately come up with arguments to support them. Seemingly vague statements like "Oh, I hate the Foo webserver, it's so flaky" or "Bar is a really annoying language to work in" or "the Baz API just doesn't feel very user-friendly" are actually the distilled product of many hard-won experiences that the speaker may not be able to call to mind individually. I observed this last night at Larry Wall's Perl 6 talk, specifically the Q&A: most of the time he came up with great examples, but sometimes he just had to talk about his own language in broad, esthetic terms. And I'm no Larry Wall, but today I found myself in a similar position when advising a client about which of two virtualization platforms to use. Of course, this applies to people speaking from experience and not from prejudice; you have to know how good someone's judgment is in order to transitively-trust them to have good opinions. That's one soft skill that gives you a major edge in technology. But, contrary to stereotype, most technical people are good at making these kinds of meta-judgments.

Tuesday, June 9, 2015

Baidu Bot Blues

A client complained of mysterious connection spikes. "We think we're getting hit by bots, but we're not sure," they said; "Can you look into it?"

So I edited postgresql.conf, turned on log_connections and log_disconnections, did a 'pg_ctl -D MY_DATA_DIR reload', and waited (in my case I also set log_min_duration_statment = 0, but I don't recommend that unless you can watch to make sure the logs aren't filling the log partition — and make sure you're not going to hammer a slow log device).

A while later, I did

ls -t data/pg_log/*.csv | head -5 | xargs grep -c connection

(That says "List log files in order from most recent to least recent, take the five most recent, and count the lines in each file containing 'connection'.")

I got lucky. A spike was obviously in progress; the current log file showed a count two orders of magnitude higher than the other log files'. Since the problem was in progress, no post mortem was required; I just mailed the client and asked "Do your web logs show bot traffic right now?"

The client mailed right back saying, "Yes, we see a bunch of connections from Baidu's web crawler; it's going nuts hitting our site."

Baidu is a Chinese search engine. It turns out its crawling bot is well-known for being aggressive. As that link explains, the bot doesn't respect Crawl-delay in robots.txt (that's annoying but unsurprising, since Crawl-delay is a non-standard extension).

So if this happens to you, you have these options:

  1. Register at Baidu's site (Chinese-only) to request less-aggressive crawling.
  2. Try to tune PostgreSQL connection pooling to handle the onslaught.
  3. Block Baidu entirely. This is an extreme option — not advised if you do business in China, or if you have Chinese-speaking customers elsewhere who are reaching you through Baidu. The latter requires analytics to figure out.

Friday, May 8, 2015

syslog Considered Dangerous

Edit 2015-10-01: there's a similar problem on RDS, where logs are written to the same partition as data, which can cause problems if you don't have enough provisioned IOPS (or disk space).

A client of ours was suffering mysterious, intermittent lock pileups on a large table. To diagnose the problem, I ran our lock-logging scripts.

Querying the resulting log_transaction_locks table, I saw lots of extend locks and RowExclusiveLocks piling up behind normal UPDATEs (and INSERTs) — there were even extend locks waiting on other extend locks. This is pretty unusual; you'd expect that when one UPDATE completed, the extend lock would succeed, PostgreSQL would quickly extend the relation (in this case, the big table's TOAST relation), and queries would keep running smoothly. These were all single-row writes; they should have taken that long. However, a complicating factor was that a couple of columns in that big table were giant TEXT columns storing JSON strings. Was that the issue? The problem was unclear, and we needed more info.

At this point my colleague Christophe Pettus made the excellent suggestion to do 'strace -p' on one of the locked processes. The dump revealed lots of sendto() calls with giant strings, which he immediately recognized as syslog messages based on their formatting.

Now, this was a problem. syslog writes to /var, and /var was hanging off the root partition, which was an EBS volume (without provisioned IOPS). So each basically the entire SQL of each query was being written across the network, including those giant JSON strings, and the query itself wouldn't terminate till the write was fsync()'ed. In the meantime, extend locks and other UPDATEs piled up behind it. The PostgreSQL data directory was on a big, fast software RAID across SSD instance storage, but performance was being bottlenecked by that single EBS volume.

The client elected to turn off log_lock_waits, and the problem went away. In addition, Christophe suggested prepending a minus sign to the filename in /etc/rsyslog.d/postgres.conf, like so, which tells syslog not to flush between each log statement:

local0.* -/var/log/postgres.log

I thought that was neat: a potential fix to a huge performance problem, in one character(!)

Lessons learned:

1. If your log_destination contains syslog, be careful about filesystem layout. (If you stick to csvlog, this normally isn't a problem, because csvlogs, by default, go in pg_log/ under the PostgreSQL data dir.)

2. Logging info about a performance problem can sometimes make it worse. In this case, we decided to diagnose the pre-existing lock pileups by setting log_lock_waits = on and deadlock_timeout = 200. But, because of the syslog issue, those very settings exacerbated the lock problems. This was necessary to catch the problem, but it's something to keep in mind if you decide to use our lock-logging scripts.