2017 job H2 3rd post

1025
data uhaul job finished really really fast
hallelujah


1027
experimental result not so good
but no clue

manager saw the search team's post about their new group result
and wanted me to do some query
during query, we found the filtering on search table provided by the search team was too strong
this could cause our training model and prediction pipeline to work weak
hallelujah


do my duty to fix some old pipeline first and then discover improvements?
maybe
you should honor God's words and show responsibility


1030
Hallelujah
peer talked to me and 'accidentally' reminded me of some impact factor again

Hallelujah
m's talk made me discovered that one filtering was too harsh


1031
Hallelujah
in some query, for simplicity and avoid cold start problem, i aggregate 2w+ term each time run the query.
now i met scalability/computation speed problem for my new pipeline
A peer messaged me to fix my pipeline to leave space for others
this also solves my new pipeline's issue.
Amazing


I was trying to talk to yi about my finding
but he just left early
this is good because i wasn't actually ready to talk to him, didn't realize this


1104
through data analysis in the morning, discovered a hard-to-notice bug
Hallelujah!

amazing things happen again and again
peer found an issue, which is also an issue to me. just hidden
Hallelujah!

talked about expectation and impact vs scope with m's manager
now the expectation is more like a good company's expectation. less about end result
Hallelujah!

anyone is helping me on things
Hallelujah!


1116
heard about classification work from a casual talk
this might lead to an important project in my next half
just leave a record here


1129
during debugging one training model command, i didn't think of using softmax is an issue, just thought the issue is from starting transformation on the group side.
i accidentally launched a command with softmax, no group tranformation, previously working training data, and found out the training still fail of low loss.
then realized softmax is an issue.
if it is really the issue, Hallelujah!

2017 H2 job 2

0901
i really long search history, and it appears that the table that i need had 180 days of retention, which is more than the normal limit of 90 days. hallelujah! this is amazing!


0903
a hive query setting missed one ;
so hard to debug, but Lord just gave me the thought to fix it

in a dataswarm job, some tasks just looked normal, but i just wanted to check them
from reading, i found out the some bugs



0905
Lord gave me the idea to check training data statistics, group by hot search term
this discovered the training data problem.
Should be a big leap

new expectation was passed down. it looks looser requirement on me than previous talked about (?)
during discussion, God help me clarify some requirements

I was about to ask m if the training data quality is good (I thought they were bad)
during discussion, found out the queries were not right
re-run, but take long to finish; promised I will arrive home for dinner, so I have to go home first
then found out the data quality is good/ok
God helped me escape one wrong statement


0906
Lord sent the peer again to me to talk to me.
he reminded me no need of so many training samples which is crucial to fast iteration, as reminded by m
also told me the pos:neg ratio that I should try.
What is happening is amazing. Hallelujah!


0908
some fblearner jobs fail because of out of memory
then the peer made recommendations and introduced one of his friend
the friend is very helpful in suggesting fix


0909
when I was about to update some complex configuration into the framework, i found the ability to switch between the config i want and previous failing config is already there

when i was trying to kick off one job, i overlooked one param
God made me see another bug which helps me see this bug

0911
while i was checking job status, found one task filed
it turned out to be a task that i don't want; but it reminds me i had another bug
amazing! hallelujah!

met one weird error in hive transformation script, which is not easy to debug and only one guy is supporting it
however, once again, the peer proactively reached out to me and told me he has met the same error before and how he overcame it.
hallelujah!


0915
这一周，因为周一晚大鲲摔倒，我陷入极度软弱中
怨恨神，和老婆打架
很想停掉给姐妹的奉献，但是网站上就是找不到login的地方
这样再祷告，也感觉圣灵不理睬

借着老婆的关爱，和神在灵修中的感动，
终于加强了信心


0916
one query just always fail
through more debugging, i found more bugs which are very easy to be ignored
many of previous jobs were not right because of this


0918
not sure if from god: late for one job
but turned out m is pto today


0921
even if yesterday's evaluation shows weird result, today's evaluation on other terms shows pretty good result!
主做事有定时
感谢赞美主！


0924
i called to zhijie about his team
he told me i happen to call at the right time, because from next week, he would work in a new team
and he also told me that current manager has the right to reject you from transferring
this is very likely to be why i am schedule to change team now: you need to do some good work first, otherwise he would have stronger hate/bias


0925
some previous experiments that looked neutral turned out to be good gain


0926
这两天一直听到这样讲道，让我们不要急于成果，用自己的方法做事，而是要效法大卫，等候神
今天早起还是有点害怕，因为mid term review要到了，开始想象review不好的结果
今天早起听到的灵修就是不要惧怕，因耶稣而得胜，在讲道和诗歌中很有感受
然后之后的工作就有不错的发现和思路

tea also helped
in the morning, i drank the half bottle tea that i deliberately left for me to drink this morning.

current model turned to be pretty good. Hallelujah!


0927
i had some very stupid ideas. God held me off discussing it to m. He just made me not willing to talk 'right now'. I felt need to use restroom and so I went. Then in the restroom, I realized how stupid my original idea was, and came up with some reasonable idea


1002
praise the Lord! I just recall there is one thing i need to check: right table for original term embed
i used 1w, but i need 1m. so this would hurt final result as fewer terms will be transformed
i just realized this and fixed it. this is so obscure to debug

debugging a weird problem
when you calm down and believe there is always enough grace from the Lord, I find out what's wrong


1003
this morning, i saw m posted in one on one about mid term expectation. i started to worry and get nervous a bit. I couldn't find anything about miracle or God's work in your lives.
my lesson is to learn to have joy no matter what circumstances
meanwhile God's promise never changes; it doesn't have to be repeated today; we should always have faith in him

knn doesn't allow using string as id, must be int
using hive to find knn met some trouble
so we have to map term -> id and use knn
previously Yi has done a mapping table from term->id. originally it was for a very weird usage: map term to id and train model. but it just gives a solution for our case above

Yi reminded me of checking term's transformed embed quality before running knn which is long
it really revealed some bugs!

repo was messed up
i happen to see one command that can help partically resolve the issue
and it turned out that my clean up work previously helps this issue as well --  not so many files to manually copy


1004
discovered a bug that cause searched name leak into training data
but the result still looks good. God's blessing



1009
last friday, the other manager rejected me
i was so frustrated
saturday, grace wasn't nice to me
and i was totally angry
sunday evening, i had headache from video and tim's talking and grace's talking

it was such an upset weekend, i was so depressed

so i went upstairs and prayed

this is the miracle, as soon as i prayed, i felt going back to God and felt much better

Hallelujah! God doesn't just change our lives, but more importantly change our hearts

today, (monday), I can still praise God and live in peace because I know my life is in his hands, and no matter what happens, God's wisdom is way higher than my wisdom

I got to know from zhijie that the reason why the other manager rejected me is:
there is a guy doing hackmonth who might be a better fit
so it just happened that someone checked the team when i was going to join. such coincidence.
but i actually feel more relieved because i know God's hand is in charge

today i also discovered that even if search->group prediction result is good, the experiment is just negative. i'm still in the progress of checking what went wrong, but so far, prediction is correctly loaded to gysj, and prediction for myself is accurate. from scuba analysis, it looks like this candidate source just really has weaker performance than other candidate sources.
if it is true that we produce good prediction result but users just don't like search, then it must be God's management as well. Even if the result looks bad, but we are in stronger peace since he is in charge.


1010
today is the true day for miracle
i knew we will talk about mid-term review
i was expecting they kick me out because project progress doesn't meet q3 goal, and received so much help from Yi
but it turned out that they are ok with progress and independency on my side
still they have concern about decision making and problem solving
those concern is also what i'm not confident in either, so i should really fix them
but i believe in God who would help me run through the difficulties and give me power
Hallelujah!!!


1013
peer's work make recommendation much better, which makes my work harder to beat it
thanks to the holy spirit and God's work in my life, I was thinking: no matter how hard the environment is, my work is in God's hand
then amazing thing happened:
i was just trying to have a chat or tease myself with the peer: your new candidate source made me think my candidate source had something wrong
then m heard it and told me he allowed me to compare my work against 'prod' before the peer's work
it is not totally weird thing, but God made it happen. if not God's work to let me chat and soften m's heart, how could this happen?


1016
recently my work has lots of issues
seed score bad
local gysj can't start
knn doesn't work
...
God has helped me find ways through


1019
think about what you have done recently
improve system performance
this is what m asked from you after review & this is also e5 expectation
it turned out that you really have the chance to work on them -- there is really problem that you can solve. also not too hard/new, not about training network, but system performance
God has paved the way for you. Hallelujah!


1020
I will run into a knn scalability issue, but I was totally not aware of it
A peer asked me about scalability of her knn job, then I realized I need to handle it as well
So I can do some early preparation to avoid after 20 hours, meet the issue and have no way in solving it
God's amazing blessing. Hallelujah!

2017 H2 job

0804
covisit offline is a project that was given to me. i didn't initiate the idea or fight for it.
the result for the first few hours looks very good


0807
during discussion, i found out my experiment was not comparing against the right control group. so it is likely that all 'gains' shown in experiment are fake.
i felt upset but thought: Lord has his time for my food
then thinking twice on the experiment, i am comparing against the right control group, which means all the gains are real.
Hallelujah!
在这个过程中，我想到offline 结果比不过online 也可能发生。所以并不是我们做的东西一定会有成果。你应当为每天的所得感谢耶和华。


0808
when i want to compute avg seed score per seed type, lots of facility is there. and there is a convenient place to keep the avg score map in local variable without the need of time-consuming new field definition

i was having trouble in grammar to populate a map in c++
but Lord just has a very easy way to use map, even for 'key not exist' case


0809
at honda service, wifi not working -> restarting computer -> lost connection to devvm where 128dim training job was running --> too hard to find command to rerun (just for backup) --> go back and check dataswarm job about what command to run --> discover the job i want actually finished even the other job timed out & didn't finish within 96 hrs
totally amazing

while checking review tool, discovered a bug that would decrease experiment performance

while i felt tired and unproductive and network got lost, I realized an issue with my experiment set up, which surely kills the experiment.
totally amazing

0811
page friend talked about using candidate source to boost high seed score
totally an idea given to me

peer shared recently liked page contextual gysj unit strong performance
and reminded me seed score may be useful
this also reminds me i can include this into broader 'recent interest' experiment

0815
during one on one, i introduced i worked on increasing recent engage seeds. y asked me reason, i said according to model analysis, those are strong features. y asked me how i analyzed, i was well prepared. i just came up with a reason that i wasn't sure: those feature scores have good importance. then y asked me to check those features and they are really important.

training jobs on hosts. i was not careful enough to clean up finished job's large input file and output file. peer reminded me. and when fixing a problem, i was given the chance to change my code to clean up the large files

while debugging a training job problem, from the misleading output, there is no way to figure out why the job fails. an idea to check stdout just came to my mind. from there, i found training was successful, it was outputting embedding that failed.


0816
once again, i wasn't sure/didn't test about one empty string handling
and m is right behind me asking if this would work
once again testing demo shows i was right

these days, i found myself not showing a good witness in multiple events:
occupying too many test segments, when asked, didn't return some to peers
delaying fixing some bugs, until upper management got noticed of it


0822
finally received note: meet expectation before end of Sep, or termination
苦难没有过去，甚至更严重
回想是不是6月时换个组，结局会不一样，抱怨神为什么不让我换组
如果换了组，就没有后面经历神，也没有灵命的改变

感谢神，今天也没有太难过，神加添给我力量在苦难中有喜乐，有顺服

不知道到底应该怎么办，祷告求神指引
没什么听见。神并不欠我们一个答案，更多时候要求我们安静等候。
尝试和yh谈谈，和j谈谈，看看有什么别的组合适

0823
今天和j聊过以后，彻底不能换组
要么坚持不懈努力达到标准，要么现在找工作

思想了一下，还是坚持不懈，不要着急采用人的方法，要安静等候神，看神的安排
等候神才有机会经历神的恩典



1. 在哪里工作，多好的公司，其实在神的国里，不重要，重要的是与神的关系和永恒的生命
2. 生命中的很多东西，像工作，神都有美好的计划，哪怕在当下人看起来不好
3. 神之前给我感动，会带我走出困境，如果包括九月的危机，那神一定会带我走过
虽然环境看起来很险恶，虽然要求很苛刻，但是在神没有难成的事，神也是信实的


0824
miracle once again: peer came to me and share low epoch also works, which greatly reduce my training time
don't put your focus in guessing if God will keep you in FB, our confidence should not be on those 'signals'; we should also not to value jobs so much.
our confidence come from Bible and what we know of God; our value should be in God's kingdom and words

an idea just came to my mind that I had better use search term embedding as search term, in order to find term internal similarity

0825
miracle once again
i didn't realize: search term may be a phrase, which contains whitespaces
it is very likely that model would treat different words in the term as separate token and we won't be able to fetch the whole term's embedding
a peer just came to me and shared his idea: use _ to concatenate
even if this may loose some information from common words, but hopefully the model can capture the relationship between the term, words in group content and group_id

even if there is a command option to achieve the result without replacing whitespaces with _, i totally forgot the command option. after talking to the peer, he reminded me of that

without those discovery, it would take super long to realize what the problem is

in the evening, the manager showed me a new approach
even if all above work need to be aborted, we still have good learning and more importantly, the new approach is more promising and can fulfill the expectation


0826
it seems much of the work needed for the new approach is already done by others. otherwise it would take really long to get something right working
amazing


0828
peer mentioned a candidate source that show what you viewed group again because he worries that candidate source already covered those groups that appear in search result but not converted
it turns out that candidate source is 'non_member_view', so it means the user was in the mall
it turns out that training search term to group still makes sense


0830
this week, my faith was in a very weak state, frightened, uncertain, worried...

around 7:30pm, received a ubn ticket
asked in the chat group and moved out back home
i really shouldn't do this. i should stay until the problem is fixed
during trip home, i also need to add gas to vehicle, which further delayed
i was thinking if asked where i was, i need to admit i went back home and shouldn't lie
this would means m will be further unhappy about me
and in this danger period this would be really bad
but i still need to admit because Lord looks over everything
it turned out m were not asking me where i was and there wasn't a serious issue
hallelujah. he is blessing

0831
the data scientist happened to show an analysis result related to a research i heard about from engage team. so I can share my learning with the team. give others an impression of involvement


0901
last night i started to doubt if Lord wants me to wait for him in this crisis
i was so doubtful that i started to think about positions in other companies
i thought about praying and listening to Lord's word again, but i thought it is very unlikely to hear anything
but in the morning, after i parked my vehicle in P&R to wait for shuttle, as soon as I started to pray, I immediately cried out. lots of tears. At that moment, I knew Lord was talking to me.
How amazing his Grace is!
I am such an ordinary person and sinful one, who did lots of things Lord is not happy with
But he still reached out to me and cared about me.
He doesn't owe me any answers, yet he talked to me and let me know his will
Hallelujah! Hallelujah! Hallelujah!

Lord blessed me to fix the bug with weird exception message quickly

had a discussion with m
had some disagreement
so i felt scared that he would give me bad scoring, also angry that he still appeared arrogant
back to home and heart g and mother had disagreement as well
it reminds me of being humble and listening
don't get angry when other people talk to you with strong opinion

2017 H1 job crisis continued 3

0705
one of the project requires running long ML training job.
and during it, jobs always fail requiring larger memory machine.
all of those happened during vacation time, especially when m is on vacation.
so we won't run into blame that training is not fast enough


0706
I totally forgot some of my work done in H1 and missed them in writing self review
God let me see one peer's comment to manager; it is a normal comment, but just attracted my attention; it reminds me of my work and reminds me of checking my quip notes about H1's work

group recommendation based on recent interest can bring good benefit, which paved the way for some of our 2017 H2 projects.

As I'm writing my PSC, I am able to describe all the challenges I tackled and engineering excellence I showed. This can possibly benefit review result and not be put into PIP. we can come back and verify this


0707
when analyzing result of one model, i queried the wrong partition, which would give me a wrong signal that the model is really really bad.


0711
my code always has a mistake that caused training file for different partitions stepped on each other
but lord made it just right: i tested with the right(normal, not reduced) volume of data
otherwise i would made a mistake

the previous idea that Lord gave me (recent interest) turned out to be very good for experiment result. even increased Meaningful People which is really hard to increase
Lord's grace is unlimited


0712
有一个training task我怎么也提高不了performance
今天上午y取消会议，我不用报告结果
我向主呼求我需要您的帮助
马上发现我之前training后忘了重新制作embedding


during evening, i found the new training job has pretty good cosine similarity. top 50 all > 0.9

0713
investigate 一个issue
结果发现data retention太短，而且cold storage也没设定
结果发现有一个table accidentally设定90天retention，正好包含我们要的数据
只有感谢神


0714
在神的保守下，发现pipeline有issue，也比较了ocis 发现的group更容易被candidate source选中被filter多次
这个任务完成了

常告诉s一些东西，也锻炼了我带junior，如果是神的特意安排，真的要感谢神


0719
之前有一个hive2file花好长时间run，使得training iteration特别慢
今天有同事主动告诉我如何降低时间，使得training time缩短很多
这样可以快速iterate

险些搞出bug
但是我想无论如何神都掌权，神都保守
结果turn out是verification时的失误

tonight, i got upset about spending too much time helping grace and why she can't be as efficient as other moms. but when i sit in front of my desk, i discovered the whole company's dataswarm pipeline are stopped because of sev...
don't know exactly why, but i should definitely continue helping grace and don't complain about it.
as for job, do my best when i have time to work, and rely on Lord for the result.


0720
talking with peer reminded me of checking file size
it turned out that i was using empty file for smoke testing
it is totally Lord's blessing and grace

when i was trying to add feature definition for seed score, i found it has been there. just not used
this great speeds up my code understanding and development cycle


when debugging why some seed not showing up on nebula side, discovered one bug on www side


0724
when backfilling table, i thought one upstream table was upfilled already
peer stopped me and asked, which reminded me to go back and check
otherwise more time will be wasted


0727
with alcohol, tired, at night
drove on 84 bridge, 50-60 mph
about to change lane to the right
didn't see a car on the right lane as its lights were too weak
right after I started to change, for some reason, I just got more alert to the right lane. So I paid more attention to the right and discovered it.
Thank you Lord for helping me out of this disaster

0728
I was blocked outside the building at eleven pm. A facility guy came out for home and opened the door for me

0731
during validation, discovered something weird (one page id is missing)
felt small and wanted to ignore
thank Lord for giving me the determination to find anything weird, even small
drilled down to root cause and found the bug that hurts ML performance badly

2017 H1 job crisis continued 2

0613
I need to use a table and the column format/type is exactly what I need
this saves tons of time of developing conversion
and there is existing query that I can borrow from, no need to write and fail and try on my own

during one on one, before i ask, i was told how to ensure certain impact
(ensure good amount of experiments that make sense)

one bad bug discovered by peer.
even if it won't affect result much, but not healthy


0614
there was a strong desire erupted in my heart that i should add one conservative treatment in one of the previous experiment
it turns out that only this treatment performs well


0615
in early morning, found metrics of early days not going up for my experiments
while walking to the office, i was in deep regret and sadness
god let the elevator's next stops disappear and still let it rise to the floors needed
don't know exactly what god is trying to say, but that's a reminder to me that god is in charge

i came up with this thought:
relying god for results and walking with god doesn't mean we can be laid back and wait for result, like what i was between march and may
you still need to work hard

0616
bug in previous experiment config about score type ( 0 to 1 or counter) emerge
during discussion, peer point out an important issue: the meaning of knn score. i may have used it in a wrong way
有个实验，需要netego的人给allocate segments，从老板到员工都不理我。也许是神预备机会让我好好思考这个出错的实验，才有机会从peer那里得到改正

你应当禁食祷告，而且奉献20%


0618
same knn score, i neglected the final stage score is already converted by previous people's code
God blessed me to discover this fact
and let me undo my conversion


0620
recent 1:1 with m's m and m's m's m are canceled for their pto or no reason
maybe this is to avoid me continuing complaining instead of looking at my own issue

while testing, discover one place that help me understand code more

a pretty hard to discover bug is noticed & found root cause of

早起发现metric大涨
下午发现看反了
虽然没有收获但是依然感恩  依然相信 依然有盼望
still i got a bit upset, then i saw this https://www.facebook.com/MartinezchurchofChrist/videos/1571368619548421/ on FB
God train us to depend on him, no matter in easy life or hard life


0621
find all experiments share same problem: high bidding & low impression
peer helped find out the one root cause for ALL of them


0622
all experiments show good early results
我是何等卑微，神竟然顾念我
神是信实的，让我们等待，时候到了就真的有恩惠
神是真神，是活神，有恩典，有大能

我们自己实在没什么可骄傲，自豪的，也不要因为成就而快乐
因为这个成就是来自于神的，我们要因为神的恩典而快乐，因经历神而快乐
好想可以因为pressure稍微降低一些而快乐
但是我们因神的喜乐应该超过眼前短暂的metric的喜乐

无论顺境逆境，我们都注目耶稣，相信神的美好带领
在逆境中，不要越想越怕，依靠神走出逆境
在顺境中，不要只因为顺利欢喜，要反复思考是神的美好恩典

for one training job, i jump to solution before we discovered noise/training time is an issue
y pointed out. at the beginning i wasn't happy. and i felt he was picky on me again.
but i recalled that i should consider my own problem first.
then after a shower, when i read the email again, i found his attitude is quite gentle and helpful

0623
一个项目是模仿page做covisit training
自己看错了人家的training data产生路线
神使我心里不安，催促我去和有经验的confirm一下training data产生方法，有经验的发现不对
这样才促使我去和author查究竟training data怎么回事， 然后改正了错误

0624
用fasttext train covisit，神提醒我应该过滤只有一个group的corpus

现在回想起2月份老板给我换项目，当时我根本不具备独自做feed ranking的能力，对各种ML algorithm 和AML的算法知之甚少。如果自己做下去，即使没有feed的人分享结果，根本做不下去


0625
我花更多时间在团契上，在孩子上，在妻子上
这本是我应当做的
神就降下恩典，让我training model顺利，产生embedding顺利而且神速
连上传照片都神速
记住，这恩典不是我配得的，不是我顺从神的报酬，是白白的恩典


0626
when comparing offline and online model training result, God has helped me realize:
before we use improvements below, maybe we want to maximize offline model performance with pure model approaches, which are unique to offline model, in order to learn if offline can really beat online

Lord has prepared GYSJ logging knowledge to me

during discussion of UCP, Feed engineer has very friendly introduced new tech two tower to us, this may bring gain

i heard from peer talk that reminds me that i should create internationalization of wording to my previous experiments


0627
i would like to find a threshold
run a couple of queries, but can't go through grammar
i have a voice inside asking is this right?
then discover the threshold i want to use and the way i use it is wrong!

manager want me to check if i was using knn in a right way
it turned out i was doing right
if it was not, it would be a big quality issue, for several other projects that involve knn as well

from the fact that i didn't use model wisely and i didn't think about how to train model dedicately, Lord has informed me there is lots of space that i need to grow. Hallelujah!


0628
we are not certain if recent seed will bring gain, because lot of times you think some work should bring gain, but it just didn't. random group + visited group is an example, I thought removing them should bring gain as rec from them don't look good to me; but it turned out that treatment was performing worse than others with those seeds

when we don't receive power from God, we feel tired easily and are poor of ideas/methods.
Even if we don't feel power from God on something, it is just our feeling: we always receive gift and wisdom from God, we just don't see it.
Also, we need to wait for God's power patiently. Appraise him and remind ourselves our blessings come from him.


0629
previously heard peer's suggestions to use candidate list for knn
now we met a painful problem that knn jobs ran too long and got killed
so his suggestions happen to be useful

in the evening, when I saw newly trained model had very weird and bad performance, my faith totally crashed. i couldn't accept it and i complained a lot to God. what a shame
God quickly reminded me it was due to kmean cluster N too large, and after i adjust it to a right number, the result looks pretty good
God forgave our sins(weak in faith...) to this degree


0630
i tried to pray on my shuttle to work, but i got easily distracted and my thought flied away by itself as well. and I was even thinking wildly & crazily as usual, having bad thoughts on the shuttle.
today i can probably check the result of the experiment launched last week. i was nervous to check because i thought my faith and prayer wasn't good enough. and if i checked, the result may look bad.
"you can never be good enough to face God"
God didn't punish me because i was not strong enough in faith and was not good enough. His grace still pour on me.
saying thank you is not enough. We raise all appraisal and glory to God!

I think what you need is to quiet down and pray again, fully commit yourself to God

those ideas come from God. you received them by yourself or from others
the work was filled with power from God. you can do them with good result and fast speed
the result was given by God. metrics growth from those experiments

1) We should realize this good experiment is God's Grace
2) Don't consider you are doing great by yourself. So no pride
3) We should be happy that God is true and alive and caring, and we can walk with God, and we have peace from him. Not so much about your experiment and you can pass this crisis
4) Continue your wait for God's plan on your crisis

understanding CNN and fasttext, previously can't understand with a whole day, now with peer's material learned from ML course, two hours are good for me


i was pretty nervous about not being able to show experiment result to yi as i see next week is the deadline of psc and he is pto
but the result of psc is totally controlled by God
showing or not showing the result doesn't matter
keep in mind who is the lord of the world and be strong in faith

2017 H1 job situation (continued)

discover experiment shows good result. B very happy
meaningful join increased 4.6%
note: conversion didn't increase at all. likely a miracle

recall: God let B talk about why positive and negative samples have different activeness. only then i realized positive samples should be filtered with meaningfulness as negative samples

privacy issue not fired at least as of 0605


你的所有神经质要求
比如：“必须走在这块砖上才能有好结果”
和各种迷信
都是拜假神



0606
Y和我说很危险，六七月必须有一些deliver

然后我的task：ranking discover category by personal timespent in the category做的就很顺利很快
12小时可以开始实验，这是前所未有的
我之前不是那么promising的project：launch movie打下了基础
耶和华的恩典早已预备，人不一定能察觉

我之前想的recent interest（page group）比较复杂。需要自己train double helix去predict page 和 group的关系。但是和yi谈完才知道我们已经有page to group covisit了，只要做好seed就可以用。极大的加速了项目


0607
昨天才做完rank discover category
今天就收到邮件别人也马上开始做
如果神没有赐恩典让我昨天迅速完成，今天这个项目就被拿走了


学习如何boost gysj value model里的source
也比较顺利


0608
以前做项目都是用几周时间，现在是几天

我在工作中觉得很有力量，工作long hour，little rest，也不觉得疲倦
也有智慧，之前意识不到的，之前想不到的idea，现在都有了

计算user这两天的 engaged page／group
有一些小case的大bug非常难发现
但是都在测试中显现出来了

连review tool都可以在desktop上run
归功于fb代码的风格，什么都可以在自己的机器上test；
也归功于神给智慧，让我知道当我需要自己的review tool时该如何开始；
也是神的恩典，让review tool在我的devserver上可以成功run；
前两点难道不是神的恩典吗？
改正：
我的想法根本错了
应该在php里加seed
感谢神的恩典让我的错误被peer更正
补充：
但是神没有让我的时间白花，
第二天就用到这天的知识debug，结果是非常confident的debug
如若不然，就只能问别人，压力大

涨metrics不是一件容易的事
看到增长，你要知道有神的恩典
其实我们有什么得着不是来自神的恩典哪？

0609
昨天花了不少时间做出的dataswarm pipeline今天可以很少修改用到新的project
神的准备超出我们所求所想
而且这个新项目本身，我没有想到的idea，神就通过y给我了

I was sure about what some fields mean; even those fields have explanation comments!

one query that exceeds memory limit in adhoc presto
doesn't exceed the limit in pipeline
so their limits are different. 感谢神



0610
因为老婆逛商场时间久，和老婆有争执
神借着这件事问我有没有把工作看得远远高于家庭
虽然表面上说工作是为了家庭，但其实内心深处是为了自己的意愿和面子
而且没有appreciate老婆一直以来的支持，虽然有时候动作慢，要求多，但是已经比别的老婆支持多了，欠老婆一个谢谢

从周五evening一直找到周日马上睡觉的bug，终于找到了
已经说自己不行了，各种方法都找不到
神最后还是给了灵感
这是不配的恩典，和白白得来的恩典

2017 End of May

Sunday morning information from Eastside 传道
work can also be God's plan on you
you should obey God's words in your work, and be a good example of Christian at work


3:00am, just recalled I need to lead bible study in coming Sat.
Afraid to forget preparing for it, but don't want to disturb sleeping. struggled...
Still opened up notebook and took note
Next morning, even if sleep was affected, but still very energetic

God takes care of you if you care about serving him first

2017 H1 crisis

Amid the sadness you can hardly meet 'meet-all' expectation, there are frustration generated from long-running tasks.
sometimes because of typo
sometimes because of careless
sometimes because of discovering new cases/learning

but I prayed hard. and sooner or later, he always helped

2017 H1 performance crisis

around mid mar, heard about:
gap between e5 expectation and reality
won't meet all goals this half

what our attitude should be:
Lord gave us the job; let me decide on what will happen, in performance review or job security
believe in Him that he will take care of your needs, financial...
think about what skills/knowledge you lack of

如何在苦难中依然喜乐，谢恩?

在苦难中，把重担交托给神，相信神的掌管，我们就有平安，就有喜乐。如果我们交托，我们的状况就是神自己的事

为什么要感谢神？

神是宇宙的主宰，创造我们的生命
神是救赎主，让我们有永恒的生命；赚得全世界，失去生命，有什么益处？
神也给我们很多保护，哪怕我们并不察觉

如何看待苦难？

任何事情神不允许不会临到，你们岂不比麻雀更珍贵吗

神不会给我们承受不了的苦难

我们本是一无所有，我们的一切都是神白白赐予的

神往往借着苦难，操练我们的品格，带给我们信心

神让我们经历苦难往往是神要把最好的东西赐给我们，也许当时看不到




神为我们舍独生爱子，还有什么不能舍呢

group highlight notification suddenly have impact

group highlight notification didn't have positive result at the beginning
at that time i missed offering for a few months
then I pleasantly made offering
without any changes, group highlight notification started to have positive results