Teh Xiggeh

There’s been quite a lot of talk about an error message discovered on a Google repository server, which has been confirmed as real by Matt Cutts. Here’s the error message in question:

pacemaker-alarm-delay-in-ms-overall-sum 2341989
pacemaker-alarm-delay-in-ms-total-count 7776761
cpu-utilization 1.28
cpu-speed 2800000000
timedout-queries_total 14227
num-docinfo_total 10680907
avg-latency-ms_total 3545152552
num-docinfo_total 10680907
num-docinfo-disk_total 2200918
queries_total 1229799558
e_supplemental=150000 –pagerank_cutoff_decrease_per_round=100 –pagerank_cutoff_increase_per_round=500 –parents=12,13,14,15,16,17,18,19,20,21,22,23 –pass_country_to_leaves –phil_max_doc_activation=0.5 –port_base=32311 –production –rewrite_noncompositional_compounds –rpc_resolve_unreachable_servers –scale_prvec4_to_prvec –sections_to_retrieve=body+url+compactanchors –servlets=ascorer –supplemental_tier_section=body+url+compactanchors –threaded_logging –nouse_compressed_urls –use_domain_match –nouse_experimental_indyrank –use_experimental_spamscore –use_gwd –use_query_classifier –use_spamscore –using_borg”

How revealing is that! Unfortunately Google say they’ve put procedures in place to stop it happening again, but I’m determined to enjoy it while it lasts.

Here’s my take on the whole deal:

pacemaker-alarm: I don’t know what this is, but it sounds like a system Google have in place to keep everything up and running. It’s an alarm, so it has a trigger. And it’s a pacemaker, so it may be triggered to prevent request timeouts. And if these numbers are right, it looks like it takes 0.301ms to trigger, and triggers on 0.6% of queries.

cpu-utilization: I assume this is a number in the same format as your standard *nix load readout. Now it could be an average over 1 minute, 5 minutes, 15 minutes, or something random Google considers important, but whatever the timescale there are 1.28 processes queued for processing on average.

cpu-speed: 2800000000 works out to be 2.8GHz - could this be a Pentium 4 with the 533MHz FSB? Or maybe Intel Xeon 2.8GHz. It’s not likely to be AMD (yet).

timedout-queries: Well it looks like queries can timeout, so maybe that’s not what the pacemaker alarm is for. But on this server 14,227 queries have timed out, which works out to be 0.0012% of total queries.

num-docinfo-total: Most likely the total number of documents stored on this particular box. If we do some number crunching with other values in the message, it looks like the average document size is 4.85KiBi. Makes sense to me, and I assume Google are still compressing documents in the repository.

avg-latency-ms_total: Could this be the average latency per query? In that case it works out to be 2.88ms per query on average.

queries_total: Fucking hell! 1.2 billion queries, presumably on that box alone.

Now all that stuff above seems to be debug values returned from the server. The next lot seems to be settings on the box;

pagerank_cutoff: Increase and decrease per round? Now we now Google needs to go through 40-60 iterations of the PageRank algorithms to get vaguely accurate figures, and documents would need to be pulled from the repository (which is what this box does). This could relate to the maximum increase/decrease in PR per iteration. It would certainly prevent drastic variations per iteration. What’s interesting are the values - “100″ and “500″. Now we see PR values 1-10 (which we know is just a fluffy number which is almost meaningless). Google patents and research papers refer to the Internet having a total PR of 1. So could these values be percentages?

parents: Every good server needs a parent, and every good system needs a backup. It appears this box has 12 sequentially-numbered parent servers, probably to assign workloads and retrieve documents.

port_base: Well I tried a port scan (sorry Google), but I really didn’t expect to find anything. Could servers in the GooglePlexi communicate in the 32000 port range?

production: Obviously, as we’re using it. It looks like Google can switch servers between testing and production at the click of a mouse.

rewrite_noncompositional_compounds: Non-Compositional Compounds (NCC) are phrases such as “hot dog” or “cold turkey” that make absolutely no sense when split up. When a computer is expected to understand the meanings of a document, and find related key terms, NCCs need to be extracted and treated differently.

scale_prvec4_to_prvec: I’m still working on the math on this one. Vectors scare me and I spent too much time on IRC at school.

sections_to_retrieve: Appears to be a list of document elements to return, either when a user requests a cached document, or when the indexer requests documents for scoring.

supplemental_tier_section: As above, but for supplemental documents?

servlets=ascorer: I don’t know what the ‘ascorer’ is, but it’s live. What begins with ‘a’ that Google would want to score? Hah, what doesn’t Google want to score

threaded_logging: That’s some serious logging going on, although I’d probably do the same.

nouse_experimental_indyrank: I can’t think what “IndyRank” might be, what it would rank or how. I want to know though

use_experimental_spamscore: No surprises there. The “bad data push” (uh huh) caused a helluva ruckus with spam in the results, and the problem is slowly going away. This appears to be our knight in shining armour.

use_gwd: Google Web Directory?

use_query_classifier: I believe this is related to Google’s OneBox results (e.g. health, stocks, companies, and so forth)

use_spamscore: Obviously this is the older method of calculating spammer pages. Didn’t work so great now, did it?

phil_max_doc_activation: Maximum document activation? Not sure on this one. I looked up some Phils at Google. We have a Phil Winterbottom who’s published some papers on Plan 9 from Bell Labs - a distributed system build from terminals, CPU servers and file servers, but that doesn’t fit.

And now I’ve finished writing this, I found another good interpretation of this error message over at Stuntdubl. It’s interesting how he’s taken the values to be related to the cached document on the server, and I assumed they were values related to the server itself.

What are your thoughts? Am I tapping away at the vague truth, or am I on the wrong track completely?

And can I have a job at Google Ireland please?